System and method for predicting future lung cancer risk

Abstract

Risk prediction models are trained and deployed to analyze images, such as computed tomography scans, to predict the future risk of lung cancer for one or more subjects. Individual risk prediction models are trained separately on nodule-specific and non-nodule-specific features, such that each risk prediction model can predict the future risk of lung cancer for different time periods (e.g., 1 year, 3 years, or 5 years). Such risk prediction models are useful for developing preventive therapies for lung cancer by enabling the enrichment of clinical trials.

Description

[Technical field]

[0001] (CROSS REFERENCE TO RELATED APPLICATIONS) This application claims the benefit of and priority to U.S. Provisional Patent Application No. 62 / 962,614, filed January 17, 2020, the entire disclosure of which is incorporated herein by reference in its entirety for all purposes. [Background technology]

[0002] Lung cancer most commonly begins with the development of lung nodules. The nodules may be cancerous or be benign overgrowths of tissue that subsequently undergo malignant transformation. To recognize this process and standardize medical management, the radiology community has developed a scoring system to stage nodules found on CT scans to assess the likelihood of the nodule becoming cancerous. This staging system, called the Lung CT Screening, Reporting and Data System (Lung-RADS), is based on the size of the nodule, the rate at which the nodule grows, and the appearance of the nodule. In general, the larger the nodule, the more rapidly it is growing, or the more irregular its appearance, the more likely it is to become cancerous. Summary of the Invention [Problem to be solved by the invention]

[0003] However, in many scenarios, a patient's lung nodules may remain undetected for a period of time, or, if detected, may already indicate an advanced stage of cancer. Therefore, there is a need for early prediction of lung cancer risk in patients, even before the development of one or more lung nodules. [Means for solving the problem]

[0004] Embodiments of the invention disclosed herein include implementing a risk prediction model to analyze images (e.g., CT scans) to predict future risk of lung cancer. The risk prediction model analyzes features extracted from the images, such as nodule-specific features and non-nodule-specific features. In various embodiments, including non-nodule-specific features in the risk prediction model has the advantage that it allows the risk prediction model to generate future risk of cancer predictions for subjects who do not yet have lung nodules. For example, by analyzing non-nodule-specific features, examples of which include lung parenchymal features and / or body composition features, the risk prediction model can be implemented for early detection of lung cancer before the development of lung nodules. In various embodiments, the risk prediction model can be implemented to predict whether a subject is likely to develop cancer in the next year. In various embodiments, the risk prediction model can be implemented to predict whether a subject is likely to develop cancer in the next 3 years. In various embodiments, the risk prediction model can be implemented to predict whether a subject is likely to develop cancer in the next 5 years. In various embodiments, the risk prediction model can be implemented to predict whether a subject is likely to develop cancer within M time periods. In various embodiments, M is any of 6 months, 1 year, 1.5 years, 2 years, 2.5 years, 3 years, 3.5 years, 4 years, 4.5 years, 5 years, 5.5 years, 6 years, 6.5 years, 7 years, 7.5 years, 8 years, 8.5 years, 9 years, 9.5 years, 10 years, 10.5 years, 11 years, 11.5 years, 12 years, 12.5 years, 13 years, 13.5 years, 14 years, 14.5 years, 15 years, 15.5 years, 16 years, 16.5 years, 17 years, 17.5 years, 18 years, 18.5 years, 19 years, 19.5 years, or 20 years.

[0005] Risk prediction models are useful for developing preventive therapies for lung cancer by enabling clinical trial enrichment and ultimately clinical treatment. For example, clinical trials will be more efficient and viable because event rates (future lung cancer) will be above background, thereby allowing enrollment of smaller cohort sizes, reducing the number of patients that need to be treated to demonstrate that the treatment prevents lung cancer. As another example, knowing the future risk of cancer for a subject allows for tailoring of clinical care to the subject. Subjects at high risk of cancer can be monitored and / or screened with increased frequency. Furthermore, subjects at high risk of cancer can be informed and / or counseled to reduce their risk (e.g., counseled to make lifestyle changes such as quitting smoking). As another example, knowing who is most likely to develop lung cancer because of certain risks to cancer treatment will impact risk / benefit decision analysis and reimbursement around the implementation of treatment in individuals.

[0006] Disclosed herein is a method for predicting one or more future risks of lung cancer in a subject, the method comprising: obtaining one or more images captured from a subject at a single time point; extracting features from the one or more obtained images, where the extracted features include at least non-nodule specific features, where the non-nodule specific features include one or both of lung parenchymal features or body composition features; and applying one or more trained risk prediction models to predict one or more future risks of lung cancer for the subject by analyzing the features extracted from the one or more obtained images. In various embodiments, predicting one or more future risks of lung cancer for the subject comprises applying an M-year risk prediction model to predict whether the subject is likely to develop lung cancer within M years, the M-year risk prediction model includes nodule-specific features and non-nodule-specific features, and more than 50% of the top three extracted features with the highest feature importance values ​​are non-nodule-specific features; the lung parenchymal features include one or more of: percentage of low attenuation areas in the lung, percentage of high attenuation areas in the lung, ratio of low or high attenuation areas in upper lung zones compared to lower lung zones, percentage of lung occupied by any of normal tissue, centrilobular emphysema, centrilobular nodules, ground glass, honeycomb, linear scars, nodular, reticular, subpleural lines, other emphysema, or cysts; and the body composition features include one or more of pectoral muscle cross-sectional area, pectoral muscle lean cross-sectional area, or subcutaneous fat cross-sectional area.

[0007] In various embodiments, the lung parenchyma features include one or more of densitometry measurements of the lung parenchyma or local histogram measurements of the lung parenchyma. In various embodiments, the densitometry measurements of the lung parenchyma include one or more of the following: percentage of low attenuation areas of the lung, percentage of high attenuation areas of the lung, and ratio of low or high attenuation areas in upper lung zones compared to lower lung zones. In various embodiments, the local histogram measurements of the lung parenchyma include one or more of the following: percentage of the lung occupied by normal tissue, centrilobular emphysema, centrilobular nodular, ground glass, honeycomb, linear scar, nodular, reticular, subpleural line, other emphysema, or cysts. In various embodiments, the body composition features include one or more of pectoral muscle cross-sectional area, pectoral muscle lean cross-sectional area, or subcutaneous fat cross-sectional area.

[0008] In various embodiments, the extracted features further include nodule-specific features. In various embodiments, the nodule-specific features include one or more of nodule-specific attenuation, nodule margin description, nodule size, nodule shape, nodule texture, nodule diameter, Lung-RADS score, or radiomics features. In various embodiments, the radiomics features include one or more of first-order statistics, 3D shape-based features, 2D shape-based features, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, adjacent graytone difference matrix, and gray-level dependency matrix. In various embodiments, the radiomics features are extracted from the transformed image by applying a filter, such as a wavelet filter or a Gaussian filter. Thus, any of first-order statistics, 3D shape-based features, 2D shape-based features, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, adjacent graytone difference matrix, and gray-level dependency matrix can be extracted from the wavelet transformed image or the Gaussian transformed image. In various embodiments, the nodule-specific features are extracted from a radiologist report. In various embodiments, the nodule-specific features are extracted computationally by implementing image analysis algorithms.

[0009] In various embodiments, predicting one or more future risks of lung cancer for the subject includes applying a 5-year risk prediction model to predict whether the subject is likely to develop lung cancer within 5 years. In various embodiments, the 5-year risk prediction model includes nodule-specific features and non-nodule-specific features, the non-nodule-specific features having a higher feature importance value than the nodule-specific features. In various embodiments, more than 50% of the top 3 extracted features with the highest feature importance values ​​are non-nodule-specific features. In various embodiments, more than 50% of the top 5 extracted features with the highest feature importance values ​​are non-nodule-specific features. In various embodiments, more than 50% of the top 10 extracted features with the highest feature importance values ​​are non-nodule-specific features. In various embodiments, determining that the subject is a candidate includes obtaining a classification of the subject in one of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, or Lung-RADS 4B. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.74. In various embodiments, the 5-year risk prediction model achieves at least a 5.1-fold increase in cumulative incidence compared to the background rate.

[0010] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as any of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, or Lung-RADS 4A. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.72. In various embodiments, the 5-year risk prediction model achieves at least a 3.9-fold increase in cumulative incidence compared to background rates.

[0011] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1, Lung-RADS 2, or Lung-RADS 3. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.67. In various embodiments, the 5-year risk prediction model achieves at least a 3.7-fold increase in cumulative incidence compared to background rates.

[0012] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1 or Lung-RADS 2. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.65. In various embodiments, the 5-year risk prediction model achieves at least a 1.9-fold increase in cumulative incidence compared to background rates.

[0013] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as Lung-RADS 1. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.57. In various embodiments, the 5-year risk prediction model achieves at least a 1.9-fold increase in cumulative incidence compared to background rates.

[0014] In various embodiments, predicting one or more future risks of lung cancer in the subject comprises applying a 3-year risk prediction model to predict whether the subject is likely to develop lung cancer within 3 years. In various embodiments, the 3-year risk prediction model comprises nodule-specific features and non-nodule-specific features, the non-nodule-specific features having higher feature importance values ​​than the nodule-specific features. In various embodiments, more than 50% of the top 5 extracted features with the highest feature importance values ​​are non-nodule-specific features. In various embodiments, more than 50% of the top 10 extracted features with the highest feature importance values ​​are non-nodule-specific features.

[0015] In various embodiments, determining that the subject is a candidate comprises obtaining a classification of the subject in one of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, or Lung-RADS 4B. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.79. In various embodiments, the 3-year risk prediction model achieves at least a 6.3-fold increase in cumulative incidence compared to background rates.

[0016] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as any of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, or Lung-RADS 4A. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.72. In various embodiments, the 3-year risk prediction model achieves at least a 5.1-fold increase in cumulative incidence compared to background rates.

[0017] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1, Lung-RADS 2, or Lung-RADS 3. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.72. In various embodiments, the 3-year risk prediction model achieves at least a 5.7-fold increase in cumulative incidence compared to background rates. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.676.

[0018] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1 or Lung-RADS 2. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.65. In various embodiments, the 3-year risk prediction model achieves at least a 3-fold increase in cumulative incidence compared to the background rate. In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as Lung-RADS 1. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.60. In various embodiments, the 3-year risk prediction model achieves at least a 2.3-fold increase in cumulative incidence compared to the background rate.

[0019] In various embodiments, predicting one or more future risks of lung cancer in a subject comprises applying a one-year risk prediction model to predict whether the subject is likely to develop lung cancer within one year. In various embodiments, the one-year risk prediction model comprises a nodule-specific feature and a non-nodule-specific feature, the non-nodule-specific feature having a higher feature importance value than the nodule-specific feature. In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1 or Lung-RADS 2. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.72. In various embodiments, the one-year risk prediction model achieves at least a 5.5-fold increase in cumulative incidence compared to the background rate. In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as Lung-RADS 1. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.57.

[0020] In various embodiments, the one-year risk prediction model includes nodule-specific features and non-nodule-specific features, the nodule-specific features having a higher feature importance value than the non-nodule-specific features. In various embodiments, more than 50% of the top three extracted features with the highest feature importance values ​​are nodule-specific features. In various embodiments, determining that the subject is a candidate includes obtaining a classification of the subject in one of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, or Lung-RADS 4B. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.90. In various embodiments, the one-year risk prediction model achieves at least an 11-fold increase in cumulative incidence compared to background rates.

[0021] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as any of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, or Lung-RADS 4A. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.81. In various embodiments, the one-year risk prediction model achieves at least a 7.6-fold increase in cumulative incidence compared to background rates.

[0022] In various embodiments, more than 50% of the top 5 extracted features with the highest feature importance values ​​are nodule-specific features. In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1, Lung-RADS 2, or Lung-RADS 3. In various embodiments, the 1-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.82. In various embodiments, the 1-year risk prediction model achieves at least an 8.6-fold increase in cumulative incidence compared to background rates.

[0023] In various embodiments, predicting one or more future risks of lung cancer in a subject comprises applying a plurality of risk prediction models to predict whether the subject is likely to develop lung cancer within N different time periods. In various embodiments, at least one of the N different time periods is any one of 6 months, 1 year, 1.5 years, 2 years, 2.5 years, 3 years, 3.5 years, 4 years, 4.5 years, 5 years, 5.5 years, 6 years, 6.5 years, 7 years, 7.5 years, 8 years, 8.5 years, 9 years, 9.5 years, 10 years, 10.5 years, 11 years, 11.5 years, 12 years, 12.5 years, 13 years, 13.5 years, 14 years, 14.5 years, 15 years, 15.5 years, 16 years, 16.5 years, 17 years, 17.5 years, 18 years, 18.5 years, 19 years, 19.5 years, or 20 years. In various embodiments, N is 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, or 20 different periods.

[0024] In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1-4B prediction model trained to predict future risk of lung cancer using training images captured from a training cohort of training individuals classified as any of Lung-RADS 1-4B. In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1-4A model trained to predict future risk of lung cancer using training images captured from a training cohort of training individuals classified as any of Lung-RADS 1-4A. In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1-3 model for predicting future risk of lung cancer using training images captured from a training cohort of training individuals classified as any of Lung-RADS 1-3. In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1-2 model that is trained to predict future risk of lung cancer using training images captured from a training cohort of training individuals classified as either Lung-RADS 1 or 2. In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1 model that is trained to predict future risk of lung cancer using training images captured from a training cohort of training individuals classified as Lung-RADS 1. In various embodiments, a majority of the training individuals in the training cohort have previously been classified as Lung-RADS 1.

[0025] In various embodiments, the one or more images are computed tomography (CT) images or X-ray images. In various embodiments, the one or more images are chest CT images or chest X-ray images. In various embodiments, the risk prediction model is trained using training images from the National Lung Screening Trial (NLST). In various embodiments, the methods disclosed herein include, prior to predicting one or more future risks of lung cancer for a subject, obtaining nodule-specific features corresponding to the subject, and determining that the subject is a candidate for future risk prediction based on the nodule-specific features.

[0026] In various embodiments, determining that the subject is a candidate comprises determining that the subject does not have lung cancer or is at low risk for developing lung cancer, hi various embodiments, determining that the subject is at low risk for developing lung cancer comprises determining that the subject does not have a nodule based on the nodule-specific features.

[0027] In various embodiments, the lung cancer is either non-small cell lung cancer or small cell lung cancer. In various embodiments, the lung cancer is either adenocarcinoma or squamous cell carcinoma. In various embodiments, the methods disclosed herein further include selecting a clinical response for the subject based on the predicted future risk of lung cancer. In various embodiments, selecting a clinical response for the subject includes selecting an intervention for treating the subject. In various embodiments, selecting an intervention includes selecting a therapeutic agent for administration to the subject. In various embodiments, the selected therapeutic agent is administered prophylactically to the subject to delay or prevent the onset of lung cancer. In various embodiments, the clinical response includes providing advice to the subject to modify the subject's behavior. In various embodiments, the clinical response includes increasing the frequency of follow-up of the subject. In various embodiments, the clinical response includes performing or scheduling to perform additional risk prediction tests to confirm the predicted future risk of lung cancer. In various embodiments, one or more of the trained risk prediction models are one of a random forest model or a gradient boosting model.

[0028] Further disclosed herein is a non-transitory computer readable medium for predicting one or more future risks of lung cancer in a subject, the non-transitory computer readable medium including instructions that, when executed by a processor, cause the processor to obtain one or more images captured from a subject at a single time point; extract features from the one or more obtained images, the extracted features including at least non-nodule specific features, the non-nodule specific features including one or both of lung parenchymal features or body composition features; and apply one or more trained risk prediction models to predict one or more future risks of lung cancer for the subject by analyzing the features extracted from the one or more obtained images. In various embodiments, the instructions causing the processor to predict one or more future risks of lung cancer for the subject, when executed by the processor, further include causing the processor to apply an M-year risk prediction model to predict whether the subject is likely to develop lung cancer within M years, the M-year risk prediction model including nodule-specific features and non-nodule-specific features, and more than 50% of the top three extracted features with the highest feature importance values ​​are non-nodule-specific features; the lung parenchymal features include one or more of: percentage of low attenuation areas in the lung, percentage of high attenuation areas in the lung, a ratio of low or high attenuation areas in the upper lung zone compared to the lower lung zone, percentage of the lung occupied by any of normal tissue, centrilobular emphysema, centrilobular nodules, ground glass, honeycomb, linear scars, nodular, reticular, subpleural lines, other emphysema, or cysts; and the body composition features include one or more of pectoral muscle cross-sectional area, pectoral muscle lean cross-sectional area, or subcutaneous fat cross-sectional area.

[0029] In various embodiments, the lung parenchyma features include one or more of densitometry measurements of the lung parenchyma or local histogram measurements of the lung parenchyma. In various embodiments, the densitometry measurements of the lung parenchyma include one or more of the following: percentage of low attenuation areas of the lung, percentage of high attenuation areas of the lung, and ratio of low or high attenuation areas in upper lung zones compared to lower lung zones. In various embodiments, the local histogram measurements of the lung parenchyma include one or more of the following: percentage of the lung occupied by normal tissue, centrilobular emphysema, centrilobular nodule, ground glass, honeycomb, linear scar, nodular, reticular, subpleural line, other emphysema, or cysts. In various embodiments, the body composition features include one or more of pectoral muscle cross-sectional area, pectoral muscle lean cross-sectional area, or subcutaneous fat cross-sectional area.

[0030] In various embodiments, the extracted features further include nodule-specific features. In various embodiments, the nodule-specific features include one or more of nodule-specific attenuation, nodule margin description, nodule size, nodule shape, nodule texture, nodule diameter, Lung-RADS score, or radiomics features. In various embodiments, the radiomics features include one or more of first-order statistics, 3D shape-based features, 2D shape-based features, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, adjacent graytone difference matrix, and gray-level dependency matrix. In various embodiments, the radiomics features are extracted from the transformed image by applying a filter, such as a wavelet filter or a Gaussian filter. Thus, any of first-order statistics, 3D shape-based features, 2D shape-based features, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, adjacent graytone difference matrix, and gray-level dependency matrix can be extracted from the wavelet transformed image or the Gaussian transformed image. In various embodiments, the nodule-specific features are extracted from a radiologist report. In various embodiments, the nodule-specific features are extracted computationally by implementing image analysis algorithms.

[0031] In various embodiments, the instructions that cause the processor to predict one or more future risks of lung cancer for the subject, when executed by the processor, further include instructions that cause the processor to apply a 5-year risk prediction model to predict whether the subject is likely to develop lung cancer within 5 years. In various embodiments, the 5-year risk prediction model includes nodule-specific features and non-nodule-specific features, the non-nodule-specific features having higher feature importance values ​​than the nodule-specific features. In various embodiments, more than 50% of the top 3 extracted features with the highest feature importance values ​​are non-nodule-specific features. In various embodiments, more than 50% of the top 5 extracted features with the highest feature importance values ​​are non-nodule-specific features. In various embodiments, more than 50% of the top 10 extracted features with the highest feature importance values ​​are non-nodule-specific features.

[0032] In various embodiments, determining that the subject is a candidate comprises obtaining a classification of the subject in one of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, or Lung-RADS 4B. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.74. In various embodiments, the 5-year risk prediction model achieves at least a 5.1-fold increase in cumulative incidence compared to background rates.

[0033] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as any of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, or Lung-RADS 4A. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.72. In various embodiments, the 5-year risk prediction model achieves at least a 3.9-fold increase in cumulative incidence compared to background rates.

[0034] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1, Lung-RADS 2, or Lung-RADS 3. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.67. In various embodiments, the 5-year risk prediction model achieves at least a 3.7-fold increase in cumulative incidence compared to background rates.

[0035] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1 or Lung-RADS 2. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.65. In various embodiments, the 5-year risk prediction model achieves at least a 1.9-fold increase in cumulative incidence compared to background rates.

[0036] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as Lung-RADS 1. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.57. In various embodiments, the 5-year risk prediction model achieves at least a 1.9-fold increase in cumulative incidence compared to background rates.

[0037] In various embodiments, the instructions that cause the processor to predict one or more future risks of lung cancer for the subject, when executed by the processor, further include instructions that cause the processor to apply a 3-year risk prediction model to predict whether the subject is likely to develop lung cancer within 3 years. In various embodiments, the 3-year risk prediction model includes nodule-specific features and non-nodule-specific features, the non-nodule-specific features having higher feature importance values ​​than the nodule-specific features. In various embodiments, more than 50% of the top 5 extracted features with the highest feature importance values ​​are non-nodule-specific features. In various embodiments, more than 50% of the top 10 extracted features with the highest feature importance values ​​are non-nodule-specific features.

[0038] In various embodiments, determining that the subject is a candidate comprises obtaining a classification of the subject in one of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, or Lung-RADS 4B. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.79. In various embodiments, the 3-year risk prediction model achieves at least a 6.3-fold increase in cumulative incidence compared to background rates.

[0039] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as any of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, or Lung-RADS 4A. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.72. In various embodiments, the 3-year risk prediction model achieves at least a 5.1-fold increase in cumulative incidence compared to background rates.

[0040] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1, Lung-RADS 2, or Lung-RADS 3. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.72. In various embodiments, the 3-year risk prediction model achieves at least a 5.7-fold increase in cumulative incidence compared to background rates. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.676.

[0041] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1 or Lung-RADS 2. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.65. In various embodiments, the 3-year risk prediction model achieves at least a 3-fold increase in cumulative incidence compared to background rates.

[0042] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as Lung-RADS 1. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.60. In various embodiments, the 3-year risk prediction model achieves at least a 2.3-fold increase in cumulative incidence compared to background rates.

[0043] In various embodiments, the instructions causing the processor to predict one or more future risks of lung cancer for the subject, when executed by the processor, further comprise instructions causing the processor to apply a one-year risk prediction model to predict whether the subject is likely to develop lung cancer within one year. In various embodiments, the one-year risk prediction model comprises nodule-specific features and non-nodule-specific features, the non-nodule-specific features having a higher feature importance value than the nodule-specific features. In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1 or Lung-RADS 2. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.72. In various embodiments, the one-year risk prediction model achieves at least a 5.5-fold increase in cumulative incidence compared to background rates.

[0044] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as Lung-RADS 1. In various embodiments, the 1-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.57.

[0045] In various embodiments, the one-year risk prediction model includes nodule-specific features and non-nodule-specific features, and the nodule-specific features have a higher feature importance value than the non-nodule-specific features. In various embodiments, more than 50% of the top three extracted features with the highest feature importance values ​​are non-nodule-specific features. In various embodiments, determining that the subject is a candidate includes obtaining a classification of the subject in one of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, or Lung-RADS 4B. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.90. In various embodiments, the one-year risk prediction model achieves at least an 11-fold increase in cumulative incidence compared to background rates.

[0046] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as any of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, or Lung-RADS 4A. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.81. In various embodiments, the one-year risk prediction model achieves at least a 7.6-fold increase in cumulative incidence compared to background rate. In various embodiments, more than 50% of the top five extracted features with the highest feature importance values ​​are non-nodule specific features.

[0047] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1, Lung-RADS 2, or Lung-RADS 3. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.82. In various embodiments, the one-year risk prediction model achieves at least an 8.6-fold increase in cumulative incidence compared to background rates.

[0048] In various embodiments, the instructions that cause the processor to predict one or more future risks of lung cancer for the subject further include instructions that, when executed by the processor, cause the processor to apply a plurality of risk prediction models to predict whether the subject is likely to develop lung cancer within N different time periods. In various embodiments, at least one of the N different periods is any one of 6 months, 1 year, 1.5 years, 2 years, 2.5 years, 3 years, 3.5 years, 4 years, 4.5 years, 5 years, 5.5 years, 6 years, 6.5 years, 7 years, 7.5 years, 8 years, 8.5 years, 9 years, 9.5 years, 10 years, 10.5 years, 11 years, 11.5 years, 12 years, 12.5 years, 13 years, 13.5 years, 14 years, 14.5 years, 15 years, 15.5 years, 16 years, 16.5 years, 17 years, 17.5 years, 18 years, 18.5 years, 19 years, 19.5 years, or 20 years. In various embodiments, N is 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, or 20 different periods.

[0049] In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1-4B prediction model trained to predict future risk of lung cancer using training images captured from a training cohort of training individuals classified as any of Lung-RADS 1-4B. In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1-4A model trained to predict future risk of lung cancer using training images captured from a training cohort of training individuals classified as any of Lung-RADS 1-4A. In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1-3 model for predicting future risk of lung cancer using training images captured from a training cohort of training individuals classified as any of Lung-RADS 1-3. In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1-2 model that is trained to predict future risk of lung cancer using training images captured from a training cohort of training individuals classified as either Lung-RADS 1 or 2. In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1 model that is trained to predict future risk of lung cancer using training images captured from a training cohort of training individuals classified as Lung-RADS 1. In various embodiments, a majority of the training individuals in the training cohort have previously been classified as Lung-RADS 1.

[0050] In various embodiments, the one or more images are computed tomography (CT) images or X-ray images. In various embodiments, the one or more images are chest CT images or chest X-ray images. In various embodiments, the risk prediction model is trained using training images from the National Lung Screening Trial (NLST).

[0051] In various embodiments, the non-transitory computer readable medium further includes instructions that, when executed by the processor, cause the processor to obtain nodule-specific features corresponding to the subject prior to predicting one or more future risks of lung cancer for the subject, and determine that the subject is a candidate for future risk prediction based on the nodule-specific features.

[0052] In various embodiments, the instructions that cause the processor to determine that the subject is a candidate further include instructions that, when executed by the processor, cause the processor to determine that the subject does not have lung cancer or is at low risk for developing lung cancer. In various embodiments, the instructions that cause the processor to determine that the subject is at low risk for developing lung cancer further include instructions that, when executed by the processor, cause the processor to determine that the subject does not have a nodule based on the nodule-specific feature.

[0053] In various embodiments, the lung cancer is either non-small cell lung cancer or small cell lung cancer. In various embodiments, the lung cancer is either adenocarcinoma or squamous cell carcinoma. In various embodiments, the non-transitory computer-readable medium disclosed herein further comprises instructions, when executed by a processor, that cause the processor to select an intervention for treating the subject based on the predicted future risk of lung cancer. In various embodiments, the instructions, when executed by the processor, that cause the processor to select an intervention further comprise instructions, when executed by the processor, that cause the processor to select a therapeutic agent for administration to the subject. In various embodiments, the selected therapeutic agent is a prophylactic agent for delaying or preventing the onset of lung cancer. In various embodiments, one or more of the trained risk prediction models are one of a random forest model or a gradient boosting model.

[0054] Further disclosed herein is a system for predicting one or more future risks of lung cancer in a subject, the system including: an imaging device configured to capture one or more images of the subject at a single time point; and a computing device configured to perform the steps of acquiring one or more images captured from the subject at the single time point; extracting features from the one or more acquired images, where the extracted features include at least non-nodule specific features, where the non-nodule specific features include one or both of lung parenchymal features or body composition features; and applying one or more trained risk prediction models to predict one or more future risks of lung cancer for the subject by analyzing the features extracted from the one or more acquired images. In various embodiments, predicting one or more future risks of lung cancer for the subject comprises applying an M-year risk prediction model to predict whether the subject is likely to develop lung cancer within M years, the M-year risk prediction model includes nodule-specific features and non-nodule-specific features, and more than 50% of the top three extracted features with the highest feature importance values ​​are non-nodule-specific features; the lung parenchymal features include one or more of: percentage of low attenuation areas in the lung, percentage of high attenuation areas in the lung, ratio of low or high attenuation areas in upper lung zones compared to lower lung zones, percentage of lung occupied by any of normal tissue, centrilobular emphysema, centrilobular nodules, ground glass, honeycomb, linear scars, nodular, reticular, subpleural lines, other emphysema, or cysts; and the body composition features include one or more of pectoral muscle cross-sectional area, pectoral muscle lean cross-sectional area, or subcutaneous fat cross-sectional area.

[0055] In various embodiments, the lung parenchyma features include one or more of densitometry measurements of the lung parenchyma or local histogram measurements of the lung parenchyma. In various embodiments, the densitometry measurements of the lung parenchyma include one or more of the following: percentage of low attenuation areas of the lung, percentage of high attenuation areas of the lung, and ratio of low or high attenuation areas in upper lung zones compared to lower lung zones. In various embodiments, the local histogram measurements of the lung parenchyma include one or more of the following: percentage of the lung occupied by normal tissue, centrilobular emphysema, centrilobular nodules, ground glass, honeycomb, linear scars, nodular, reticular, subpleural lines, other emphysema, or cysts. In various embodiments, the body composition features include one or more of pectoral muscle cross-sectional area, pectoral muscle lean cross-sectional area, or subcutaneous fat cross-sectional area.

[0056] In various embodiments, the extracted features further include nodule-specific features. In various embodiments, the nodule-specific features include one or more of nodule-specific attenuation, nodule margin description, nodule size, nodule shape, nodule texture, nodule diameter, Lung-RADS score, or radiomics features. In various embodiments, the radiomics features include one or more of first-order statistics, 3D shape-based features, 2D shape-based features, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, adjacent graytone difference matrix, and gray-level dependency matrix. In various embodiments, the radiomics features are extracted from the transformed image by applying a filter, such as a wavelet filter or a Gaussian filter. Thus, any of first-order statistics, 3D shape-based features, 2D shape-based features, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, adjacent graytone difference matrix, and gray-level dependency matrix can be extracted from the wavelet transformed image or the Gaussian transformed image. In various embodiments, the nodule-specific features are extracted from a radiologist report. In various embodiments, the nodule-specific features are extracted computationally by implementing image analysis algorithms.

[0057] In various embodiments, predicting one or more future risks of lung cancer in the subject comprises applying a 5-year risk prediction model to predict whether the subject is likely to develop lung cancer within 5 years. In various embodiments, the 5-year risk prediction model comprises nodule-specific features and non-nodule-specific features, the non-nodule-specific features having higher feature importance values ​​than the nodule-specific features. In various embodiments, more than 50% of the top 3 extracted features with the highest feature importance values ​​are non-nodule-specific features. In various embodiments, more than 50% of the top 5 extracted features with the highest feature importance values ​​are non-nodule-specific features. In various embodiments, more than 50% of the top 10 extracted features with the highest feature importance values ​​are non-nodule-specific features.

[0058] In various embodiments, determining that the subject is a candidate comprises obtaining a classification of the subject in one of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, or Lung-RADS 4B. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.74. In various embodiments, the 5-year risk prediction model achieves at least a 5.1-fold increase in cumulative incidence compared to background rates.

[0059] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as any of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, or Lung-RADS 4A. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.72. In various embodiments, the 5-year risk prediction model achieves at least a 3.9-fold increase in cumulative incidence compared to background rates.

[0060] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1, Lung-RADS 2, or Lung-RADS 3. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.67. In various embodiments, the 5-year risk prediction model achieves at least a 3.7-fold increase in cumulative incidence compared to background rates.

[0061] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1 or Lung-RADS 2. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.65. In various embodiments, the 5-year risk prediction model achieves at least a 1.9-fold increase in cumulative incidence compared to background rates.

[0062] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as Lung-RADS 1. In various embodiments, the 5-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.57. In various embodiments, the 5-year risk prediction model achieves at least a 1.9-fold increase in cumulative incidence compared to background rates.

[0063] In various embodiments, predicting one or more future risks of lung cancer in the subject comprises applying a 3-year risk prediction model to predict whether the subject is likely to develop lung cancer within 3 years. In various embodiments, the 3-year risk prediction model comprises nodule-specific features and non-nodule-specific features, the non-nodule-specific features having higher feature importance values ​​than the nodule-specific features. In various embodiments, more than 50% of the top 5 extracted features with the highest feature importance values ​​are non-nodule-specific features. In various embodiments, more than 50% of the top 10 extracted features with the highest feature importance values ​​are non-nodule-specific features.

[0064] In various embodiments, determining that the subject is a candidate comprises obtaining a classification of the subject in one of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, or Lung-RADS 4B. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.79. In various embodiments, the 3-year risk prediction model achieves at least a 6.3-fold increase in cumulative incidence compared to background rates.

[0065] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as any of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, or Lung-RADS 4A. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.72. In various embodiments, the 3-year risk prediction model achieves at least a 5.1-fold increase in cumulative incidence compared to background rates.

[0066] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1, Lung-RADS 2, or Lung-RADS 3. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.72. In various embodiments, the 3-year risk prediction model achieves at least a 5.7-fold increase in cumulative incidence compared to background rates. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.676.

[0067] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1 or Lung-RADS 2. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.65. In various embodiments, the 3-year risk prediction model achieves at least a 3-fold increase in cumulative incidence compared to background rates.

[0068] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as Lung-RADS 1. In various embodiments, the 3-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.60. In various embodiments, the 3-year risk prediction model achieves at least a 2.3-fold increase in cumulative incidence compared to background rates.

[0069] In various embodiments, predicting one or more future risks of lung cancer in the subject comprises applying a one-year risk prediction model to predict whether the subject is likely to develop lung cancer within one year. In various embodiments, the one-year risk prediction model comprises nodule-specific features and non-nodule-specific features, and the non-nodule-specific features have higher feature importance values ​​than the nodule-specific features.

[0070] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1 or Lung-RADS 2. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.72. In various embodiments, the one-year risk prediction model achieves at least a 5.5-fold increase in cumulative incidence compared to background rates. In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as Lung-RADS 1. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.57. In various embodiments, the one-year risk prediction model includes nodule-specific features and non-nodule-specific features, the nodule-specific features having a higher feature importance value than the non-nodule-specific features. In various embodiments, more than 50% of the top three extracted features with the highest feature importance values ​​are non-nodule-specific features.

[0071] In various embodiments, determining that the subject is a candidate comprises obtaining a classification of the subject in one of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, or Lung-RADS 4B. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.90. In various embodiments, the one-year risk prediction model achieves at least an 11-fold increase in cumulative incidence compared to background rates.

[0072] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as any of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, or Lung-RADS 4A. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.81. In various embodiments, the one-year risk prediction model achieves at least a 7.6-fold increase in cumulative incidence compared to background rate. In various embodiments, more than 50% of the top five extracted features with the highest feature importance values ​​are non-nodule specific features.

[0073] In various embodiments, determining that the subject is a candidate further comprises determining that the subject is classified as either Lung-RADS 1, Lung-RADS 2, or Lung-RADS 3. In various embodiments, the one-year risk prediction model exhibits an area under the curve (AUC) value of at least 0.82. In various embodiments, the one-year risk prediction model achieves at least an 8.6-fold increase in cumulative incidence compared to background rates.

[0074] In various embodiments, predicting one or more future risks of lung cancer in a subject comprises applying a plurality of risk prediction models to predict whether the subject is likely to develop lung cancer within N different time periods. In various embodiments, at least one of the N different time periods is any one of 6 months, 1 year, 1.5 years, 2 years, 2.5 years, 3 years, 3.5 years, 4 years, 4.5 years, 5 years, 5.5 years, 6 years, 6.5 years, 7 years, 7.5 years, 8 years, 8.5 years, 9 years, 9.5 years, 10 years, 10.5 years, 11 years, 11.5 years, 12 years, 12.5 years, 13 years, 13.5 years, 14 years, 14.5 years, 15 years, 15.5 years, 16 years, 16.5 years, 17 years, 17.5 years, 18 years, 18.5 years, 19 years, 19.5 years, or 20 years. In various embodiments, N is 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, or 20 different periods.

[0075] In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1-4B prediction model trained to predict future risk of lung cancer using training images captured from a training cohort of training individuals classified as any of Lung-RADS 1-4B. In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1-4A model trained to predict future risk of lung cancer using training images captured from a training cohort of training individuals classified as any of Lung-RADS 1-4A. In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1-3 model for predicting future risk of lung cancer using training images captured from a training cohort of training individuals classified as any of Lung-RADS 1-3. In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1-2 model that is trained to predict future risk of lung cancer using training images captured from a training cohort of training individuals classified as either Lung-RADS 1 or 2. In various embodiments, one of the one or more risk prediction models is a Lung-RADS 1 model that is trained to predict future risk of lung cancer using training images captured from a training cohort of training individuals classified as Lung-RADS 1. In various embodiments, a majority of the training individuals in the training cohort have previously been classified as Lung-RADS 1.

[0076] In various embodiments, the one or more images are computed tomography (CT) images or X-ray images. In various embodiments, the one or more images are chest CT images or chest X-ray images. In various embodiments, the risk prediction model is trained using training images from the National Lung Screening Trial (NLST).

[0077] In various embodiments, the computing device is configured to, prior to predicting one or more future risks of lung cancer for the subject, obtain a nodule-specific feature corresponding to the subject and determine that the subject is a candidate for future risk prediction based on the nodule-specific feature. In various embodiments, determining that the subject is a candidate includes determining that the subject does not have lung cancer or is at low risk of developing lung cancer. In various embodiments, determining that the subject is at low risk of developing lung cancer includes determining that the subject does not have nodules based on the nodule-specific feature.

[0078] In various embodiments, the lung cancer is either non-small cell lung cancer or small cell lung cancer. In various embodiments, the lung cancer is either adenocarcinoma or squamous cell carcinoma. In various embodiments, the computing device is further configured to perform the step of selecting an intervention for treating the subject based on the predicted future risk of lung cancer. In various embodiments, selecting an intervention for treating the subject includes selecting a therapeutic agent for administration to the subject. In various embodiments, the selected therapeutic agent is a prophylactic agent for delaying or preventing the onset of lung cancer. In various embodiments, one or more of the trained risk prediction models is one of a random forest model or a gradient boosting model. [Brief description of the drawings]

[0079] These and other features, aspects and advantages of the present invention will become better understood with regard to the following description and accompanying drawings. [Figure 1A] 1 shows an overview of a system environment for determining a cancer prognosis in a subject, according to one embodiment. [Figure 1B] FIG. 1 shows a block diagram of a cancer prediction system, according to one embodiment. [Figure 2A] FIG. 1 shows an exemplary flow diagram for determining a subject's future risk of cancer for applications such as patient enrichment, according to a first embodiment. [Figure 2B]FIG. 1 shows an exemplary flow diagram for determining future risk of cancer for a subject for uses such as patient enrichment, according to a second embodiment. [Diagram 3] 1 is an exemplary flow process for determining a subject's future risk of cancer for applications such as patient enrichment, according to one embodiment. [Figure 4] 1 illustrates an exemplary computer for implementing the entities illustrated in FIGS. 1A, 1B, 2A, 2B, and 3. FIG. [Figure 5A] Shows the performance of risk prediction models to predict the probability of cancer within 1 year across patients with Lung-RADS 1-4B. [Figure 5B] Shows the performance of risk prediction models to predict the likelihood of cancer within 3 years across patients with Lung-RADS 1-4B. [Figure 5C] We show the performance of risk prediction models to predict the probability of cancer within 5 years across patients with Lung-RADS 1-4B. [Figure 5D] Three-year cumulative incidence functions are shown across Lung-RADS 1–4B patients. [Figure 6A] Shows the performance of risk prediction models to predict the probability of cancer within 1 year across Lung-RADS 1-4A patients. [Figure 6B] Shows the performance of risk prediction models to predict the likelihood of cancer within 3 years across Lung-RADS 1-4A patients. [Figure 6C] We show the performance of risk prediction models to predict the probability of cancer within 5 years across Lung-RADS 1-4A patients. [Figure 6D] Three-year cumulative incidence functions across Lung-RADS 1-4A patients are shown. [Figure 7A] Shows the performance of risk prediction models to predict the probability of cancer within 1 year across Lung-RADS 1-3 patients. [Figure 7B]We show the performance of risk prediction models to predict the likelihood of cancer within 3 years across Lung-RADS 1-3 patients. [Figure 7C] We show the performance of risk prediction models to predict the likelihood of cancer within 5 years across Lung-RADS 1-3 patients. [Figure 7D] Three-year cumulative incidence functions across Lung-RADS 1–3 patients are shown. [Figure 8A] We show the performance of risk prediction models to predict the probability of cancer within 1 year across Lung-RADS 1-2 patients. [Figure 8B] We show the performance of risk prediction models to predict the likelihood of cancer within 3 years across Lung-RADS 1-2 patients. [Figure 8C] We show the performance of risk prediction models to predict the likelihood of cancer within 5 years across Lung-RADS 1-2 patients. [Figure 8D] Three-year cumulative incidence functions across Lung-RADS 1–2 patients are shown. [Figure 9A] Shows the performance of risk prediction models to predict the probability of cancer within 1 year across Lung-RADS 1 patients. [Figure 9B] Shows the performance of risk prediction models to predict the likelihood of cancer within 3 years across Lung-RADS 1 patients. [Figure 9C] Shows the performance of risk prediction models to predict the probability of cancer within 5 years across Lung-RADS 1 patients. [Figure 9D] Three-year cumulative incidence functions across Lung-RADS 1 patients are shown. [Figure 10A] We demonstrate the performance of risk prediction models incorporating radiomics features to predict the probability of cancer within 1 year across Lung-RADS 1-4B patients. [Figure 10B] We show the performance of risk prediction models incorporating radiomics features to predict the probability of cancer within 3 years across Lung-RADS 1-4B patients. [Figure 11A]Shows the performance of risk prediction models incorporating radiomics features to predict the probability of cancer within 1 year across Lung-RADS 1-4A patients. [Figure 11B] We show the performance of risk prediction models incorporating radiomics features to predict the probability of cancer within 3 years across Lung-RADS 1-4A patients. [Figure 12A] We demonstrate the performance of risk prediction models incorporating radiomics features to predict the probability of cancer within 1 year across Lung-RADS 1-3 patients. [Figure 12B] We demonstrate the performance of risk prediction models incorporating radiomics features to predict the likelihood of cancer within 3 years across Lung-RADS 1-3 patients. [Figure 13A] We demonstrate the performance of risk prediction models incorporating radiomics features to predict the probability of cancer within 1 year across Lung-RADS 2-4B patients. [Figure 13B] We demonstrate the performance of risk prediction models incorporating radiomics features to predict the probability of cancer within 3 years across Lung-RADS 2-4B patients. [Figure 14A] Shows the performance of a risk prediction model incorporating radiomics features to predict the probability of cancer within 1 year across Lung-RADS 4A-4B patients. [Figure 14B] Shows the performance of a risk prediction model incorporating radiomics features to predict the probability of cancer within 3 years across Lung-RADS 4A-4B patients. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0080] I. Definition Terms used in the claims and specification are defined as set forth below, unless otherwise stated.

[0081] The terms "subject" or "patient" are used interchangeably and include cells, tissues, or organisms, whether in vivo, ex vivo, or in vitro, male or female, human or non-human.

[0082] The term "mammal" encompasses both humans and non-humans, and includes, but is not limited to, humans, non-human primates, canines, felines, murines, bovines, equines, and porcines.

[0083] The term "sample" or "test sample" can include a single cell or multiple cells or fragments of a cell or bodily fluid aliquot, such as a blood sample, taken from a subject by means including venipuncture, voiding, ejaculation, massage, biopsy, needle aspirate, lavage, scraping, surgical incision, or other means, or other means known in the art. Examples of bodily fluid aliquots include amniotic fluid, aqueous humor, bile, lymph, breast milk, interstitial fluid, blood, plasma, earwax, cow's fluid, chyle, chyme, female ejaculate, menses, mucus, saliva, urine, vomit, tears, vaginal fluid, sweat, serum, semen, sebum, pus, pleural effusion, cerebrospinal fluid, synovial fluid, intracellular fluid, and vitreous fluid. In various embodiments, the sample can be a biopsy of a tissue, such as a lung tumor or lung nodule.

[0084] The term "acquiring one or more images" encompasses acquiring one or more images captured from a subject or acquiring one or more images captured from a sample obtained from a subject. Acquiring one or more images can encompass performing a step of capturing one or more images from a subject or from a sample obtained from a subject. The phrase can also encompass receiving one or more images, for example, from a third party who performed a step of capturing one or more images from a subject or from a sample obtained from a subject. The one or more images can be acquired by one of skill in the art through various known methods, including being stored in a storage device. In various embodiments, "acquiring one or more images" refers to acquiring one or more images, each captured from a subject at a single time point (e.g., a single patient visit).

[0085] The term "training images" refers to images (e.g., CT images or X-ray images) captured from an individual used to train a risk prediction model, e.g., a lung cancer risk prediction model, such as a model described herein. In various embodiments, the training images are computed tomography (CT) images from a cohort built from routine clinical care of patients (e.g., from routinely screened patients). In various embodiments, the training images are computed tomography (CT) images from a cohort built from a research study (e.g., a federal / industry sponsored research study). In various embodiments, the training images are computed tomography (CT) images from the National Lung Cancer Screening Trial (NLST) cohort. In various embodiments, the training images are computed tomography (CT) images included in a custom dataset. For example, the training images can be captured from a training individual. The term "training individual" refers to an individual from whom training images are captured or otherwise obtained for use in training a risk prediction model.

[0086] The term "nodule-specific features" refers to features of a lung nodule, examples of which include nodule-specific attenuation, nodule margin description, nodule size, nodule shape, nodule texture (e.g., smooth, spiculated, etc.), nodule diameter, and Lung-RADS score. In various embodiments, the nodule-specific features are computationally extracted from the image (e.g., by implementing an image analysis algorithm). For example, the nodule-specific features can be radiomics features extracted using an image processing algorithm such as PyRadiomics. Exemplary radiomics features can include first-order statistics, 3D shape-based features, 2D shape-based features, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, adjacent graytone difference matrix, and gray-level dependency matrix. In various embodiments, the radiomics features are extracted from the transformed image by applying a filter such as a wavelet filter or a Gaussian filter. Thus, any of the following may be extracted from the wavelet transformed or Gaussian transformed images: first order statistics, 3D shape-based features, 2D shape-based features, gray level co-occurrence matrices, gray level run length matrices, gray level size zone matrices, adjacent gray tone difference matrices, and gray level dependency matrices. As used herein, "nodule specific features" are also referred to as "subjective features." In various embodiments, the nodule specific features are extracted from reports made by trained experts (e.g., radiologists) who analyze images of a subject, such as a CT scan of the subject.

[0087] The term "non-nodule specific features" refers to either lung parenchymal features or body composition features. Non-nodule specific features can include features from subjects that are not nodules. Non-nodule specific features can include features from subjects that are not associated with nodules. Lung parenchymal features can include densitometry measurements of lung parenchyma, which can include, for example, the percentage of the lung occupied by (i) low attenuation areas (LAA), defined as areas / volumes with attenuation lower than -950 Hounsfield Units (HU), and (ii) high attenuation areas (HAA), defined as areas / volumes of the lung with attenuation between -600 HU and -250 HU, and the ratio of LAA in the upper lung zone to LAA in the lower lung zone (ratio LAA). Lung parenchymal features may further include measurements of interstitial changes of the lung parenchyma, such as regional histogram measurements of the lung parenchyma, e.g., percentage of the lung occupied by normal tissue, centrilobular emphysema, centrilobular nodules, ground glass, honeycomb, linear scars, nodular, reticular, subpleural lines, other emphysema, and / or cysts. Body composition may include, for example, pectoralis major muscle cross-sectional area, pectoralis minor muscle cross-sectional area, pectoralis major muscle lean cross-sectional area, pectoralis minor muscle lean cross-sectional area, left or right pectoralis major or pectoralis minor muscle collective cross-sectional area, and subcutaneous fat cross-sectional area. As used herein, "non-nodule specific features" are also referred to as "objective features". In various embodiments, "non-nodule specific features" are computationally extracted from images, such as images captured from a subject. In various embodiments, non-nodule features computationally extracted from features may be used to construct a report, such as a radiologist report, including the non-nodule features. In various embodiments, "non-nodule specific features" do not include duration of emphysema or duration of cardiovascular disease.

[0088] The phrase "Lung-RADS X through Y" is meant to include individuals classified as Lung-RADS X, Lung-RADS Y, and any value between the numerical variables "X" and "Y." For example, Lung-RADS 1 through 3 refers to individuals classified as Lung-RADS 1, Lung-RADS 2, and Lung-RADS 3.

[0089] The phrase "Lung-RADS XY prediction model" refers to a risk prediction model that is trained using training images captured from training individuals classified into any one of Lung-RADS XY.

[0090] The phrase "M-year predictive model" refers to a risk prediction model trained to predict future risk of cancer within "M" time period. In various embodiments, "M" is any of 6 months, 1 year, 1.5 years, 2 years, 2.5 years, 3 years, 3.5 years, 4 years, 4.5 years, 5 years, 5.5 years, 6 years, 6.5 years, 7 years, 7.5 years, 8 years, 8.5 years, 9 years, 9.5 years, 10 years, 10.5 years, 11 years, 11.5 years, 12 years, 12.5 years, 13 years, 13.5 years, 14 years, 14.5 years, 15 years, 15.5 years, 16 years, 16.5 years, 17 years, 17.5 years, 18 years, 18.5 years, 19 years, 19.5 years, or 20 years. In various embodiments, "M" is measured from the time of acquisition of one or more images from the subject. Thus, a one-year prediction model refers to a risk prediction model that is trained to predict the future risk of cancer within one year from the time of acquisition of one or more images from a subject.

[0091] The phrase "M year Lung-RADS XY prediction model" refers to a risk prediction model that is 1) trained using training images captured from training individuals classified as any one of Lung-RADS X through Y, and 2) trained to predict future risk of cancer within "M" years.

[0092] The phrase "future risk of cancer" refers to the risk of a subject developing cancer within a given time period, e.g., within 1 year or 3 years from T0. In various embodiments, future risk of cancer refers to the likelihood that a subject will develop cancer within a given time period from time zero (T0). In various embodiments, future risk of cancer refers to the likelihood that a subject will develop cancer within 1 year. In various embodiments, future risk of cancer refers to the likelihood that a subject will develop cancer within 3 years. In various embodiments, future risk of cancer refers to the likelihood that a subject will develop cancer within 5 years. In various embodiments, future risk of cancer refers to the likelihood that a subject will develop cancer within at least 1 year, at least 2 years, at least 3 years, at least 4 years, at least 5 years, at least 6 years, at least 7 years, at least 8 years, at least 9 years, at least 10 years, at least 11 years, at least 12 years, at least 13 years, at least 14 years, at least 15 years, at least 16 years, at least 17 years, at least 18 years, at least 19 years, or at least 20 years. In various embodiments, the "future risk of cancer" is a binary value (e.g., 0 or 1, where 0 indicates a low likelihood of developing cancer during the time period and 1 indicates a high likelihood of developing cancer during the time period). In various embodiments, the "future risk of cancer" is a continuous value (e.g., between 0 and 1, where values ​​closer to 1 indicate a high likelihood of developing cancer during the time period).

[0093] The terms "treating", "treatment", or "therapy" of lung cancer shall mean administering a treatment to slow, stop, or reverse the progression of the cancer. In some embodiments, treating lung cancer means reversing the progression of the cancer, ideally to the point of eliminating the cancer itself. In various embodiments, "treating", "treatment", or "therapy" of lung cancer includes administering a therapeutic agent or pharmaceutical composition to a subject. Furthermore, as used herein, "treating", "treatment", or "therapy" of lung cancer further includes administering a therapeutic agent or pharmaceutical composition for preventive purposes. Prevention of cancer refers to the administration of a composition or therapeutic agent to prevent the onset, development, manifestation, progression, or recurrence of cancer, or some or all of the symptoms of lung cancer, or to reduce the likelihood of developing lung cancer.

[0094] It must be noted that, as used herein, the singular forms "a," "an," and "the" include plural references unless the context clearly dictates otherwise.

[0095] II. System Environment Overview FIG 1A illustrates an overview of a system environment for determining a cancer prediction for a subject, according to one embodiment. The system environment 100 provides a context for introducing a subject 110, an imaging device 120, and a cancer prediction system 130 for determining a cancer prediction 140 for the subject 110. Although FIG 1A illustrates one subject 110 for which a cancer prediction 140 is generated, in various embodiments, the system environment 100 includes two or more subjects such that the cancer prediction system 130 generates cancer predictions 140 for two or more subjects (e.g., a cancer prediction for each subject).

[0096] In various embodiments, the subject 110 is healthy. For example, the subject has not been previously diagnosed with cancer or is not suspected of having cancer. Thus, the method for predicting future risk of cancer described herein can be useful for early detection of cancer in healthy subjects. In certain embodiments, the type of cancer in the subject is lung cancer. Thus, the method described herein can be useful for early detection of lung cancer. In various embodiments, the subject has previously been diagnosed with cancer. In such an embodiment, the subject can be in remission, and thus, the method for predicting future risk of cancer can be useful for determining whether the subject is likely to experience a recurrence of cancer in a period of time.

[0097] In various embodiments, subjects predicted to develop cancer or experience cancer recurrence within a time period can be administered a treatment, such as a prophylactic treatment, to delay or prevent the onset or recurrence of cancer. In various embodiments, subjects predicted to develop cancer or experience cancer recurrence within a time period are selected to be enrolled in a clinical trial.

[0098] Referring to FIG. 1A, the imaging device 120 captures an image from the subject 110. In various embodiments, the imaging device 120 captures an image from a test sample obtained from the subject 110. In various embodiments, the image and / or sample may be acquired by a third party, such as a medical professional. Examples of medical professionals include doctors, emergency medical technicians, nurses, first responders, psychologists, phlebotomists, physical therapy personnel, nurse practitioners, surgeons, dentists, and any other obvious medical professionals as known to those of skill in the art. In various embodiments, the image and / or sample may be acquired at a hospital or medical clinic. In various embodiments, the image and / or sample may be captured using an imaging device, such as a mobile imaging device.

[0099] In some embodiments, imaging device 120 captures images of anatomical locations of subject 110. Exemplary anatomical locations of the subject may include lungs, thoracic cavity, kidneys, liver, pancreas, brain, stomach, intestines, hips, knees, legs, arms, and face. In various embodiments, imaging device 120 captures images of the thoracic cavity of subject 110. In various embodiments, imaging device 120 captures images of the lungs of the subject. In various embodiments, imaging device 120 captures images of the chest of the subject (e.g., chest wall including pectoral muscles). In various embodiments, imaging device 120 captures images of the thoracic cavity including the lungs of the subject. In various embodiments, imaging device 120 captures images of the thoracic cavity including the chest of the subject (e.g., chest wall including pectoral muscles). In various embodiments, imaging device 120 captures images of the thoracic cavity including both the lungs of the subject and the chest of the subject (e.g., chest wall including pectoral muscles).

[0100] In various embodiments, imaging device 120 is one of a computed tomography (CT) scanner, a magnetic resonance imaging (MRI) scanner, a positron emission tomography (PET) scanner, an X-ray scanner, or an ultrasound imaging device. In certain embodiments, imaging device 120 is a CT scanner that captures one or more images of subject 110. In certain embodiments, imaging device 120 is a CT scanner that captures one or more CT images of the chest cavity including both the subject's lungs and the subject's chest (e.g., chest wall including pectoral muscles). In certain embodiments, imaging device 120 is an X-ray scanner that captures one or more X-ray images of the chest including both the subject's lungs and the subject's chest (e.g., chest wall including pectoral muscles).

[0101] In general, cancer prediction system 130 analyzes one or more images captured of subject 110 (e.g., images taken by imaging device 120) and generates a cancer prediction 140 for subject 110. In various embodiments, cancer prediction 140 determined by cancer prediction system 130 is a predicted future risk of cancer for subject 110. For example, cancer prediction 140 is a value that indicates whether subject 110 is predicted to develop cancer within a certain period of time (e.g., within one year, within three years, or within five years) from the date an image was captured of subject 110.

[0102] In various embodiments, to generate the cancer prediction 140, the cancer prediction system 130 extracts features from one or more images and applies one or more trained risk prediction models to analyze the features of the one or more images. The trained risk prediction models predict the future risk of cancer for the subject 110 within a time period. For example, the cancer prediction system 130 can apply a risk prediction model trained to predict the future risk of cancer within three years. In various embodiments, the cancer prediction system 130 determines multiple future risks of cancer over different time periods for the subject 110 by applying multiple trained risk prediction models. For example, in addition to applying a risk prediction model trained to predict the future risk of cancer within three years, the cancer prediction system 130 further applies a second risk prediction model trained to predict the future risk of cancer within five years. The cancer prediction system 130 can apply more trained risk prediction models trained for additional time periods (e.g., 1 year, 10 years, 15 years, 20 years, etc.). Generally, risk prediction models are trained independently and are not additive (e.g., you cannot subtract cancers predicted at 1 year from those predicted at 3 years to get cancers predicted to occur at 1 year but less than 3 years).

[0103] In various embodiments, the cancer prediction 140 is an index derived from the subject's predicted future risk of cancer and identifies whether the subject 110 will be included or excluded from a patient cohort for enrolling in a clinical trial. This index is useful for enrichment purposes in clinical trials. For example, if the subject 110 is predicted to develop cancer within a certain period of time, the index identifies the subject 110 to be included in a patient cohort for enrolling in a clinical trial. As another example, if the subject 110 is not predicted to develop cancer within a certain period of time, the index identifies the subject 110 to be excluded from a patient cohort for enrolling in a clinical trial.

[0104] In various embodiments, the cancer prediction 140 may include a recommended intervention for the subject 110 based on the predicted future risk of cancer. For example, if the cancer prediction system 130 determines that the subject 110 is likely to develop cancer within one year, the cancer prediction 140 may include a recommended intervention to delay or prevent the rapid onset of cancer over the next year.

[0105] Cancer prediction system 130 may include one or more computers embodied as a computer system 400, as discussed below with respect to Figure 4. Thus, in various embodiments, the steps described with reference to cancer prediction system 130 are performed in silico.

[0106] In various embodiments, imaging device 120 and cancer prediction system 130 are used by different parties. For example, a first party operates imaging device 120 to capture one or more images from subject 110 and then provides the captured one or more images to a second party that implements cancer prediction system 130 to determine cancer prediction 140. In some embodiments, imaging device 120 and cancer prediction system 130 are used by the same party.

[0107] Reference is now made to FIG. 1B, which illustrates a block diagram showing computer logic components of a cancer prediction system 130, according to one embodiment. Here, the cancer prediction system 130 includes a feature extraction module 145, a candidate object module 150, a risk training module 155, a risk expansion module 160, an enrichment module 165, and a training data store 170. In various embodiments, the cancer prediction system 130 can be configured differently with additional or fewer modules. For example, the cancer prediction system 130 need not include the candidate object module 150. As another example, the cancer prediction system 130 need not include the risk training module 155 or the training data store 170 (as shown by the dotted lines in FIG. 1B), and instead, the risk training module 155 and the training data store 170 can be used by different systems and / or parties.

[0108] Generally, the feature extraction module 145 extracts features from images captured from the subject or training images captured from the training individuals. In various embodiments, the feature extraction module 145 extracts non-nodule specific features from the images or training images. In various embodiments, the feature extraction module 145 extracts nodule specific features from the images or training images. In various embodiments, the feature extraction module 145 extracts nodule specific features and non-nodule specific features from the images or training images. The feature extraction module 145 provides the features extracted from the training images to a risk training module 155 for training a risk prediction model. In various embodiments, the feature extraction module 145 provides the features extracted from the images captured from the subject to a risk to candidate object module 150 for identifying candidate objects. In various embodiments, the feature extraction module 145 provides the features extracted from the images captured from the subject to a risk deployment module 160 for deploying one or more trained risk prediction models.

[0109] The candidate object module 150 analyzes features extracted from one or more images captured from a subject to determine whether the subject is a candidate object for receiving a future risk prediction. This is useful for identifying a subset of patients to receive a future risk prediction. For example, it may be preferable to predict the future risk of cancer for low-risk cancer patients (e.g., patients who do not currently have cancer and / or patients who do not currently have lung nodules indicative of cancer). Thus, the candidate object module 150 can identify a subset of low-risk cancer patients for subsequent future risk prediction. In various embodiments, the candidate object module 150 analyzes both nodule-specific features and non-nodule-specific features to determine whether the subject is a candidate object. In various embodiments, the candidate object module 150 analyzes only nodule-specific features to determine whether the subject is a candidate object. In various embodiments, the candidate object module 150 need not be implemented by the cancer staging system 140. For example, in a scenario in which all subjects are analyzed for future cancer risk, all subjects are candidate objects to receive a future risk prediction.

[0110] The risk training module 155 trains a risk prediction model using training data derived from a training individual. For example, the training data includes features extracted from one or more training images captured from the training individual. In various embodiments, the risk training module 155 trains a risk prediction model that includes both nodule-specific features and non-nodule-specific features. In such embodiments, the risk expansion module 160 implements a risk prediction model that analyzes both nodule-specific features and non-nodule-specific features extracted from images acquired from a subject (e.g., subject 110 of FIG. 1A ) to determine a future risk of cancer. In various embodiments, the risk training module 155 trains a risk prediction model that includes only non-nodule-specific features. In such embodiments, the risk expansion module 160 implements a risk prediction model that analyzes only non-nodule-specific features extracted from images acquired from a subject (e.g., subject 110 of FIG. 1A ) to determine a future risk of cancer.

[0111] The risk deployment module 160 implements a risk prediction model to analyze features extracted from images acquired from a subject (e.g., subject 110 of FIG. 1A ) to determine a cancer prediction, such as a prediction of future risk of cancer for the subject 110. The training and deployment risk prediction models are described in further detail below.

[0112] The enrichment module 165 selects subjects for inclusion in a patient cohort. In various embodiments, the patient cohort is generated for enrollment in a clinical trial. For example, the enrichment module 165 generates a patient cohort (e.g., an enriched patient cohort) that includes a higher percentage of patients who will develop cancer within a certain period of time. This enriched patient cohort allows for enrollment of fewer patients in clinical trials and / or identifies people who may benefit from treatments to prevent and prevent lung cancer.

[0113] The components of the cancer prediction system 130 are described below with reference to two phases: 1) a training phase and 2) a deployment phase. More specifically, the training phase refers to the construction and training of one or more risk prediction models by the risk training module 155 based on training data, such as training images captured from training individuals (e.g., individuals known to either develop or not develop cancer within a certain time period). Thus, the models are trained using the training data such that implementation of the models by the risk deployment module 160 during the deployment phase enables prediction of a subject's (e.g., subject 110 of FIG. 1A ) future risk of cancer.

[0114] In some embodiments, the components of the cancer prediction system 130 are applied during one of the training phase and the deployment phase. For example, the risk training module 155 and the training data store 170 are applied to train the risk model during the training phase. Furthermore, the risk deployment module 160 is applied during the deployment phase. In various embodiments, the components of the cancer prediction system 130 can be executed by different parties depending on whether the components are applied during the training phase or the deployment phase. In such a scenario, the training and deployment of the predictive model are performed by different parties. For example, the risk training module 155 and the training data store 170 applied during the training phase can be used by a first party (e.g., to train the risk prediction model), and the risk deployment module 160 applied during the deployment phase can be executed by a second party (e.g., to deploy the risk prediction model).

[0115] III. Methods for predicting future risk of cancer The embodiments described herein include a method for determining a subject's future risk of cancer by applying one or more trained risk prediction models. Such a method can be performed by the cancer prediction system 130 described in FIG. 1B. Further reference is made to FIG. 2A, which shows an exemplary flow diagram 200 for determining a subject's future risk of cancer for applications such as patient enrichment, according to one embodiment.

[0116] As shown in Figure 2A, a captured image of the object 210 is obtained from an object (e.g., object 110 of Figure 1A). In various embodiments, the image of the object 210 is a CT image captured by performing a CT scan of the object. In various embodiments, the image of the object 210 is an X-ray image captured by performing an X-ray scan of the object. In various embodiments, more than one image of the object 210 is captured from the object.

[0117] In various embodiments, the feature extraction module 145 extracts features 215A of the subject. In various embodiments, the features 215A of the subject include clinical data corresponding to the subject, such as age, sex, ethnicity, smoking history, geographic location, pollution exposure, and / or family history of lung cancer. In various embodiments, the feature extraction module 145 (FIG. 1B) extracts the features 215A from the subject image 210. In various embodiments, the feature extraction module 145 implements an image analysis algorithm to extract the features 215A from the subject image 210. In various embodiments, the feature extraction module 145 implements an image analysis algorithm that includes a machine learning model trained to analyze and extract features from an image. Methods for extracting radiomics features are further described in Radiomics of Lung Nodules: A multi-institutional study of robustness and agreement of quantitative imaging features. Tomography. 2016; 2(4): 430-437 and Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012; 48(4): 441-446, each of which is incorporated by reference in its entirety.

[0118] In various embodiments, feature extraction module 145 extracts at least two features from target image 210. In various embodiments, feature extraction module 145 extracts at least three, at least four, at least five, at least six, at least seven, at least eight, at least nine, or at least ten features from target image 210. In various embodiments, feature extraction module 145 extracts at least 15, at least 20, at least 25, at least 30, at least 35, at least 40, or at least 45 features from target image 210. In various embodiments, feature extraction module 145 extracts at least 50 features from target image 210. In various embodiments, feature extraction module 145 extracts at least 100 features, at least 150 features, at least 200 features, at least 250 features, at least 300 features, at least 350 features, at least 400 features, at least 450 features, at least 500 features, at least 550 features, at least 600 features, at least 650 features, at least 700 features, at least 750 features, at least 800 features, at least 850 features, at least 900 features, at least 950 features, or at least 1000 features from target image 210. In various embodiments, feature extraction module 145 extracts between 100 features and 1000 features. In various embodiments, feature extraction module 145 extracts between 300 features and 900 features. In various embodiments, feature extraction module 145 extracts between 500 features and 1000 features.

[0119] In various embodiments, the feature extraction module 145 extracts features 215A including nodule-specific features. Nodule-specific features refer to features of a lung nodule (e.g., a lung nodule that is present or absent in the target image 210). Exemplary nodule-specific features include nodule-specific attenuation, nodule margin description, nodule size, nodule shape, nodule texture (e.g., smooth, spiculated, etc.), nodule diameter, and Lung-RADS score. In various embodiments, the nodule-specific features can be radiomics features extracted using image processing algorithms such as PyRadiomics. Exemplary radiomics features can include first-order statistics, 3D shape-based features, 2D shape-based features, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, adjacent graytone difference matrix, and gray-level dependency matrix. In various embodiments, the radiomics features are extracted from the transformed image by applying a filter such as a wavelet filter or a Gaussian filter. Thus, any of the following can be extracted from the wavelet transformed or Gaussian transformed images: first order statistics, 3D shape-based features, 2D shape-based features, grey level co-occurrence matrix, grey level run length matrix, grey level size zone matrix, adjacent grey tone difference matrix, and grey level dependency matrix.

[0120] In certain embodiments, feature extraction module 145 analyzes target image 210 and assigns a Lung-RADS score to target image 210 based on one or more of the extracted nodule-specific features. For example, based on one or more extracted nodule-specific features (e.g., radiomics features, etc.), feature extraction module 145 determines that target image 210 does not contain a lung nodule. In such a scenario, feature extraction module 145 may assign target image 210 a Lung-RADS score of 1. As another example, feature extraction module 145 analyzes target image 210 and determines that target image 210 contains a nodule based on one or more of the extracted nodule-specific features. Based on the nodule-specific features, feature extraction module 145 can then assign a corresponding Lung-RADS score (e.g., Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, Lung-RADS 4B, or Lung-RADS 4X) according to a Lung-RADS criteria, such as the current Lung-RADS criteria shown in Table 1 or future Lung-RADS criteria.

[0121] In various embodiments, the feature extraction module 145 extracts features 215A including non-nodule specific features. Non-nodule specific features refer to any of the lung parenchymal features (e.g., densitometry measurements of lung parenchyma and measurements of interstitial changes of lung parenchyma) and muscle / chest wall body composition measurements. Densitometry measurements of lung parenchyma may include, for example, the percentage of the lung occupied by (i) low attenuation areas (LAA), defined as the area / volume with attenuation lower than -950 Hounsfield Units (HU), and (ii) high attenuation areas (HAA), defined as the area / volume of the lung with attenuation between -600 HU and -250 HU, and the ratio of LAA in the upper lung zone to that in the lower lung zone. Measures of interstitial change of the lung parenchyma may further include regional histogram measures of the lung parenchyma, such as the percentage of the lung occupied by normal tissue, centrilobular emphysema, centrilobular nodular, ground glass, honeycomb, linear scar, nodular, reticular, subpleural line, other emphysema, and / or cysts. Body composition measures of musculature / chest wall may include, for example, pectoralis major cross-sectional area, pectoralis minor cross-sectional area, pectoralis major lean cross-sectional area, pectoralis minor lean cross-sectional area, left or right pectoralis major or pectoralis minor collective cross-sectional area, and subcutaneous fat cross-sectional area.

[0122] In various embodiments, the feature extraction module 145 extracts features 215A including nodule-specific features and non-nodule-specific features. In various embodiments, the feature extraction module 145 extracts features 215A including only non-nodule-specific features. In certain embodiments, the feature extraction module 145 extracts features 215A including only nodule-specific features. In various embodiments, the feature extraction module 145 obtains nodule-specific features determined by a third party. For example, the nodule-specific features may be determined by a trained expert (e.g., a radiologist) analyzing the target image 210. In various embodiments, the feature extraction module 145 extracts the nodule-specific features from a report generated by a third party. For example, the feature extraction module 145 extracts the nodule-specific features from a report generated by a trained expert (e.g., a radiologist) analyzing the target image 210.

[0123] The candidate subject module 150 (FIG. 1B) performs subject analysis 220 (shown in FIG. 2A) by analyzing the features 215A of the subject image 210. Based on the analysis, the candidate subject module 150 determines whether the subject is a candidate subject to undergo future risk prediction. In other words, the subject analysis 220 is a screening process to identify candidate subjects who qualify for future risk of cancer prediction. For example, a subject undergoing lung cancer screening who is determined to be free of common lung cancer is then evaluated for the risk of developing future incident lung cancer.

[0124] As an example, the candidate subject module 150 determines that a subject is a candidate subject for receiving future risk prediction if the subject's characteristics 215A, such as the subject's clinical data, meet one or more criteria. For example, if the subject's characteristics 215A indicate that the subject smokes above a threshold amount, the subject is considered a candidate subject for receiving future risk prediction. As another example, the candidate subject module 150 determines that a subject is a candidate subject for receiving future risk prediction if the subject is a low-risk cancer patient (e.g., a patient who does not currently have lung cancer and / or a patient who does not currently have lung nodules indicative of cancer). In various embodiments, the candidate subject module 150 determines that a subject is a non-candidate subject for receiving future risk prediction if the subject is a high-risk cancer patient (e.g., a patient who currently has lung cancer and / or a patient who has lung nodules indicative of a high risk of developing lung cancer). As shown in FIG. 2A, if the subject analysis 220 determines that the subject is a candidate subject, the subject image 210 is taken as a candidate subject image 225 (shown by a dotted line) for subsequent future risk prediction analysis. Alternatively, if the subject analysis 220 determines that the subject is a non-candidate subject, the patient does not undergo further future risk prediction analysis. As an example, an intervention for a non-candidate subject 218 can be provided to the subject. For example, a non-candidate subject can be a subject who already has lung cancer or has a lung nodule indicating that the subject has progressive lung cancer. Thus, the non-candidate subject would not need to undergo a future risk of cancer prediction and instead, an intervention for the non-candidate subject 218 can be provided, which can include a cancer biopsy and / or administration of a therapeutic agent (e.g., chemotherapy, radiation) to treat the non-candidate subject's lung cancer.

[0125] In various embodiments, the candidate subject module 150 determines that a subject is a candidate subject based on features 215A, including nodule-specific features. By way of example, the nodule-specific features may include a Lung-RADS score, such as Lung-RADS 0, Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, Lung-RADS 4B, or Lung-RADS 4X. A summary of Lung-RADS score classifications and corresponding features for lung nodules is set forth in Table 1. In various embodiments, the candidate subject module 150 determines a Lung-RADS score for a subject based on features 215A, including nodule-specific features. For example, the nodule-specific features may include one or more of nodule-specific attenuation, nodule margin description, nodule size, nodule shape, nodule texture (e.g., smooth, spiculated, etc.), nodule diameter, Lung-RADS score, and / or radiomics features such as first-order statistics, 3D shape-based features, 2D shape-based features, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, adjacent graytone difference matrix, and gray-level dependency matrix. Thus, if the nodule attenuation features indicate that the lung nodule is a solid nodule and the nodule margin and / or nodule diameter features indicate that the lung nodule is 5 mm, the candidate subject module 150 may assign a Lung-RADS score of 2 based on the criteria specified in Table 1.

[0126] In various embodiments, the candidate subject module 150 determines that a subject is a candidate subject if the subject is assigned a Lung-RADS score of 0. In various embodiments, the candidate subject module 150 determines that a subject is a candidate subject if the subject is assigned a Lung-RADS score of 1. In various embodiments, the candidate subject module 150 determines that a subject is a candidate subject if the subject is assigned a Lung-RADS score of 2. In various embodiments, the candidate subject module 150 determines that a subject is a candidate subject if the subject is assigned a Lung-RADS score of 3. In various embodiments, the candidate subject module 150 determines that a subject is a candidate subject if the subject is assigned a Lung-RADS score of 4A. In various embodiments, the candidate subject module 150 determines that a subject is a candidate subject if the subject is assigned a Lung-RADS score of 0 or 1. In various embodiments, the candidate subject module 150 determines that a subject is a candidate subject if the subject is assigned a Lung-RADS score of 0, 1, or 2. In various embodiments, candidate subject module 150 determines that a subject is a candidate subject if the subject has been assigned a Lung-RADS score of 0, 1, 2, or 3. In various embodiments, candidate subject module 150 determines that a subject is a candidate subject if the subject has been assigned a Lung-RADS score of 1, 2, or 3.

[0127] In various embodiments, candidate subject module 150 determines that a subject is a non-candidate subject if the subject is assigned a Lung-RADS score of 3. In various embodiments, candidate subject module 150 determines that a subject is a non-candidate subject if the subject is assigned a Lung-RADS score of 4A. In various embodiments, candidate subject module 150 determines that a subject is a non-candidate subject if the subject is assigned a Lung-RADS score of 4B. In various embodiments, candidate subject module 150 determines that a subject is a non-candidate subject if the subject is assigned a Lung-RADS score of 4A or 4B. In various embodiments, candidate subject module 150 determines that a subject is a non-candidate subject if the subject is assigned a Lung-RADS score of 3, 4A, or 4B.

[0128] In various embodiments, the candidate object module 150 determines that a subject is a candidate subject if the nodule-specific features indicate that the subject does not have a pulmonary nodule. For example, the nodule-specific features may include one or more of nodule attenuation, nodule margin description, or nodule diameter. Thus, if the nodule-specific features indicate that the subject image 210 does not contain a nodule (e.g., lack of attenuation, lack of margin, or near-zero or zero diameter), the candidate object module 150 determines that the subject is a candidate subject.

[0129] In various embodiments, the candidate object module 150 determines that a subject is a candidate subject even if the subject has a pulmonary nodule. For example, if the nodule-specific features indicate that the subject has a pulmonary nodule, the candidate object module 150 can further analyze the nodule-specific features to classify the nodule as a higher risk nodule or a lower risk nodule. In various embodiments, the candidate object module 150 can classify the nodule based on whether it is a solid nodule, a semi-solid nodule, or a non-solid nodule. For example, the candidate object module 150 can classify the nodule as a higher risk nodule if it is a solid nodule or a semi-solid nodule, and as a lower risk nodule if it is a non-solid nodule. In various embodiments, the candidate object module 150 determines that a subject is a candidate subject if the subject has a lower risk nodule. The candidate object module 150 can determine that a subject is a non-candidate subject if the subject has a higher risk nodule.

[0130] Returning to FIG. 2A, after the object analysis 220, the object image 210 from the candidate object is now considered as the candidate object image 225. The feature extraction module 145 extracts features 215B from the candidate object image 225. In various embodiments, the feature extraction module 145 extracts features 215B including one or both of nodule-specific features and non-nodule-specific features from the candidate object image 225. In various embodiments, the feature extraction module 145 extracts features 215A including only non-nodule-specific features. In various embodiments, the feature extraction module 145 extracts features 215A including only nodule-specific features. In various embodiments, the feature extraction module 145 obtains nodule-specific features of the candidate object image 225 that are determined by a third party. For example, the nodule-specific features may be determined by a trained expert (e.g., a radiologist) that analyzes the candidate object image 225. In various embodiments, the feature extraction module 145 extracts the nodule-specific features from a report generated by a third party. For example, the feature extraction module 145 extracts nodule-specific features from a report generated by a trained expert (eg, a radiologist) analyzing the target image 225 .

[0131] In various embodiments, one or more of features 215B are the same as one or more of features 215A. Thus, the same features do not need to be extracted again and can be reused. For example, the nodule-specific features previously extracted from target image 210 as features 215A can be the same nodule-specific features included in features 215B. In various embodiments, all of features 215B were previously extracted from target image 210, so the previously extracted features can be reused here as features 215B. For example, feature extraction module 145 may extract features 215A and 215B prior to target analysis 220. Thus, feature 215B does not need to be further extracted from candidate target image 225 and can be reused here.

[0132] In various embodiments, features 215B include additional features that were not included in features 215A. For example, features 215B include non-nodule specific features that were not previously included in features 215A. Thus, feature extraction module 145 extracts these additional features that were not previously included in features 215A.

[0133] In various embodiments, the feature extraction module 145 extracts at least two features from the candidate target images 225. In various embodiments, the feature extraction module 145 extracts at least 3, at least 4, at least 5, at least 6, at least 7, at least 8, at least 9, or at least 10 features from the candidate target images 225. In various embodiments, the feature extraction module 145 extracts at least 15, at least 20, at least 25, at least 30, at least 35, at least 40, or at least 45 features from the candidate target images 225. In various embodiments, the feature extraction module 145 extracts at least 50 features from the candidate target images 225. In various embodiments, the feature extraction module 145 extracts at least 100 features, at least 150 features, at least 200 features, at least 250 features, at least 300 features, at least 350 features, at least 400 features, at least 450 features, at least 500 features, at least 550 features, at least 600 features, at least 650 features, at least 700 features, at least 750 features, at least 800 features, at least 850 features, at least 900 features, at least 950 features, or at least 1000 features from the candidate target images 225. In various embodiments, the feature extraction module 145 extracts between 100 features and 1000 features from the candidate target images 225. In various embodiments, the feature extraction module 145 extracts between 300 features and 900 features from the candidate target images 225. In various embodiments, the feature extraction module 145 extracts between 500 and 1000 features from the candidate target images 225 .

[0134] The risk expansion module 160 (FIG. 1B) provides the extracted features 215B to a trained risk prediction model 230 (shown in FIG. 2A) to generate a risk prediction 240. In various embodiments, as shown in FIG. 2A, the risk expansion module 160 provides the extracted features 215B to multiple trained risk prediction models 230 to generate multiple risk predictions 240. In various embodiments, the risk prediction models are trained to generate future risk of cancer predictions within a time period (e.g., future risk of cancer within one year, within three years, or within five years). Thus, to generate predictions for multiple time periods, the risk expansion module 160 selects and expands different risk prediction models to analyze the extracted features 215B. For example, the risk expansion module 160 can expand a first risk prediction model trained to predict future risk of cancer within a first time period and can further expand a second risk prediction model trained to predict future risk of cancer within a second time period.

[0135] In various embodiments, the risk expansion module 160 expands 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, or 20 different risk prediction models to generate a future risk of cancer over 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, or 20 different time periods, respectively. In various embodiments, the risk expansion module 160 expands five different risk prediction models to generate a future risk of cancer over five different time periods. In various embodiments, the risk expansion module 160 expands three different risk prediction models to generate a future risk of cancer over three different time periods. For example, the risk expansion module 160 expands a 1-year risk prediction model, a 3-year risk prediction model, and a 5-year risk prediction model to generate a prediction of a future risk of cancer within 1 year, 3 years, and 5 years, respectively.

[0136] In various embodiments, each of the one or more risk prediction models 230 deployed to analyze the features 215B was previously trained on training images separated into different regions (e.g., different lung regions). For example, a first risk prediction model 230 may be trained to predict the presence of cancer in a first region of the lung, a second risk prediction model 230 may be trained to predict the presence of cancer in a second region of the lung, and a third risk prediction model 230 may be trained to predict the presence of cancer in a third region of the lung. By way of example, the different lung regions may include the upper third, middle third, and lower third of the lung by lung volume or separate lobes. Thus, the risk prediction 240 generated for the subject may be a future risk of cancer in a particular region (e.g., lung region).

[0137] In various embodiments, each of the one or more risk prediction models 230 deployed to analyze features 215B was previously trained using a cohort of training individuals that match the cohort of candidate subjects determined as a result of the subject analysis 220. In other words, the risk prediction models 230 were previously trained using training individuals that would qualify as candidate subjects if they were to undergo the subject analysis 220.

[0138] In various embodiments, one or more of the training individuals used to train the risk prediction model 230 share at least one feature with the candidate subject. As an example, if the candidate subject is classified as Lung-RADS 1, then one or more of the training individuals used to train the risk prediction model 230 were also classified as Lung-RADS 1. In various embodiments, the risk prediction model 230 may be trained using training images captured from training individuals classified as Lung-RADS 1 (referred to as a Lung-RADS 1 risk prediction model). In various embodiments, the risk prediction model 230 may be trained using training images captured from training individuals classified as Lung-RADS 1 or Lung-RADS 2 (referred to as a Lung-RADS 1-2 risk prediction model). In various embodiments, the risk prediction model 230 may be trained using training images captured from training individuals classified as Lung-RADS 1, Lung-RADS 2, or Lung-RADS 3 (referred to as a Lung-RADS 1-3 risk prediction model). In various embodiments, the risk prediction model 230 may be trained using training images captured from training individuals classified as Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, or Lung-RADS 4A (referred to as a Lung-RADS 1-4A risk prediction model). In various embodiments, the risk prediction model 230 may be trained using training images captured from training individuals classified as Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, or Lung-RADS 4B (referred to as a Lung-RADS 1-4B risk prediction model). In various embodiments, the risk prediction model 230 may be trained using training images captured from training individuals classified as Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, Lung-RADS 4B, or Lung-RADS 4X (referred to as a Lung-RADS 1-4X risk prediction model).

[0139] In a scenario where a candidate subject is classified as Lung-RADS 1 (e.g., classified via subject analysis 220 or previously classified by a third party), risk expansion module 160 expands one or more risk prediction models 230 trained using training images captured from at least one training individual also previously classified as Lung-RADS 1. Thus, risk expansion module 160 can expand one or more risk prediction models 230, each of which is either a Lung-RADS 1 prediction model, a Lung-RADS 1-2 prediction model, a Lung-RADS 1-3 prediction model, a Lung-RADS 1-4A prediction model, a Lung-RADS 1-4B prediction model, or a Lung-RADS 1-4X model. In certain embodiments, risk expansion module 160 expands a Lung-RADS 1 prediction model for the Lung-RADS 1 candidate subject. In certain embodiments, the risk expansion module 160 expands a Lung-RADS 1-2 predictive model for Lung-RADS 1 candidate subjects. In certain embodiments, the risk expansion module 160 expands a Lung-RADS 1-3 predictive model for Lung-RADS 1 candidate subjects.

[0140] In a scenario where a candidate subject is classified as Lung-RADS 2 (e.g., classified via subject analysis 220 or previously classified by a third party), risk expansion module 160 expands one or more risk prediction models 230 trained using training images captured from at least one training individual also previously classified as Lung-RADS 2. Thus, risk expansion module 160 can expand one or more risk prediction models 230, each of which is either a Lung-RADS 1-2 predictive model, a Lung-RADS 1-3 predictive model, a Lung-RADS 1-4A predictive model, a Lung-RADS 1-4B predictive model, or a Lung-RADS 1-4X model. In certain embodiments, risk expansion module 160 expands a Lung-RADS 1-2 predictive model for the Lung-RADS 2 candidate subject. In certain embodiments, risk expansion module 160 expands a Lung-RADS 1-3 predictive model for the Lung-RADS 2 candidate subject.

[0141] In a scenario where a candidate subject is classified as Lung-RADS 3 (e.g., classified via subject analysis 220 or previously classified by a third party), risk expansion module 160 expands one or more risk prediction models 230 trained using training images captured from at least one training individual that was also previously classified as Lung-RADS 3. Thus, risk expansion module 160 can expand one or more risk prediction models 230, each of which is either a Lung-RADS 1-3 prediction model, a Lung-RADS 1-4A prediction model, a Lung-RADS 1-4B prediction model, or a Lung-RADS 1-4X model.

[0142] In a scenario where a candidate subject is classified as Lung-RADS 4A (e.g., classified via subject analysis 220 or previously classified by a third party), risk expansion module 160 expands one or more risk prediction models 230 trained using training images captured from at least one training individual that was also previously classified as Lung-RADS 4A. Thus, risk expansion module 160 can expand one or more risk prediction models 230, each of which is either a Lung-RADS 1-4A prediction model, a Lung-RADS 1-4B prediction model, or a Lung-RADS 1-4X model.

[0143] In a scenario where a candidate subject is classified as Lung-RADS 4B (e.g., classified via subject analysis 220 or previously classified by a third party), risk expansion module 160 expands one or more risk prediction models 230 trained using training images captured from at least one training individual also previously classified as Lung-RADS 4B. Thus, risk expansion module 160 can expand one or more risk prediction models 230, each of which is either a Lung-RADS 1-4B prediction model or a Lung-RADS 1-4X model.

[0144] In a scenario where a candidate subject is classified as Lung-RADS 4X (e.g., classified via subject analysis 220 or previously classified by a third party), risk expansion module 160 expands one or more risk prediction models 230 trained using training images captured from at least one training individual that was also previously classified as Lung-RADS 4X. Thus, risk expansion module 160 can expand one or more risk prediction models 230, each of which is a Lung-RADS 1-4X prediction model.

[0145] In various embodiments, the risk prediction model is 1) trained using a cohort of training individuals that match the cohort of candidate subjects determined as a result of the subject analysis 220, and 2) trained to generate a future risk of cancer prediction within a time period (e.g., future risk of cancer within 1 year, 3 years, or 5 years). For example, to generate a risk prediction 240 for a candidate subject, the risk expansion module 160 applies one or more risk prediction models 230 that are 1) trained using one or more training individuals that share at least one characteristic with the candidate subject, and 2) trained to predict the future risk of cancer within a different time period.

[0146] In a scenario where a candidate subject is classified as Lung-RADS "Z" (e.g., classified via subject analysis 220 or previously classified by a third party), risk deployment module 160 deploys one or more risk prediction models 230 that are 1) trained using training images captured from at least one training individual that was also previously classified as Lung-RADS "Z" and 2) predict future risk of cancer within different time periods. As used herein, these risk prediction models are referred to as "M-year, Lung-RADS X-Y prediction models," where "M" refers to the time period and "X-Y" refers to the range of Lung-RADS scores of the training individuals. For example, a 1-year, Lung-RADS 1-3 prediction model refers to a risk prediction model trained using training individuals previously classified as Lung-RADS 1-3 and trained to predict future risk of cancer within 1 year.

[0147] In various embodiments, for a candidate subject classified as Lung-RADS 1, risk expansion module 160 expands one or more M-year, Lung-RADS X-Y predictive models, where "M" is variable while "X" and "Y" are fixed. For example, for a candidate subject classified as Lung-RADS 1, "M" can range from 1-5 years, where X=1 and Y=any value from 1-4B. As an example, if Y=3, risk expansion module 160 can expand a 1-year, Lung-RADS 1-3 predictive model, a 2-year, Lung-RADS 1-3 predictive model, a 3-year, Lung-RADS 1-3 predictive model, a 4-year, Lung-RADS 1-3 predictive model, and a 5-year, Lung-RADS 1-3 predictive model. In other embodiments, the risk expansion module 160 can expand additional risk prediction models beyond those described herein (e.g., the range of M is 1-10, 1-15, or 1-20 years, e.g., X and Y are selected differently to provide different ranges of Lung-RADS scores).

[0148] In certain embodiments in which a candidate subject is classified as Lung-RADS 1, the risk expansion module 160 expands: 1) a first risk prediction model 230 that predicts future risk of cancer within 1 year (e.g., 1 year, Lung-RADS 1-3 predictive model), 2) a second risk prediction model 230 that predicts future risk of cancer within 3 years (e.g., 3 year, Lung-RADS 1-3 predictive model), and 3) a third risk prediction model 230 that predicts future risk of cancer within 5 years (e.g., 5 year, Lung-RADS 1-3 predictive model).

[0149] In certain embodiments in which a candidate subject is classified as Lung-RADS 2, the risk expansion module 160 expands: 1) a first risk prediction model 230 that predicts the future risk of cancer within 1 year (e.g., 1 year, Lung-RADS 1-3 predictive model), 2) a second risk prediction model 230 that predicts the future risk of cancer within 3 years (e.g., 3 year, Lung-RADS 1-3 predictive model), and 3) a third risk prediction model 230 that predicts the future risk of cancer within 5 years (e.g., 5 year, Lung-RADS 1-3 predictive model).

[0150] In certain embodiments in which a candidate subject is classified as Lung-RADS 3, the risk expansion module 160 expands: 1) a first risk prediction model 230 that predicts future risk of cancer within 1 year (e.g., 1 year, Lung-RADS 1-3 predictive model), 2) a second risk prediction model 230 that predicts future risk of cancer within 3 years (e.g., 3 year, Lung-RADS 1-3 predictive model), and 3) a third risk prediction model 230 that predicts future risk of cancer within 5 years (e.g., 5 year, Lung-RADS 1-3 predictive model).

[0151] In certain embodiments in which a candidate subject is classified as Lung-RADS 4A, the risk expansion module 160 expands 1) a first risk prediction model 230 that predicts the future risk of cancer within 1 year (e.g., 1 year, Lung-RADS 1-4A predictive model), 2) a second risk prediction model 230 that predicts the future risk of cancer within 3 years (e.g., 3 year, Lung-RADS 1-4A predictive model), and 3) a third risk prediction model 230 that predicts the future risk of cancer within 5 years (e.g., 5 year, Lung-RADS 1-4A predictive model).

[0152] In certain embodiments in which a candidate subject is classified as Lung-RADS 4B, the risk expansion module 160 expands 1) a first risk prediction model 230 that predicts the future risk of cancer within 1 year (e.g., 1 year, Lung-RADS 1-4B predictive model), 2) a second risk prediction model 230 that predicts the future risk of cancer within 3 years (e.g., 3 year, Lung-RADS 1-4B predictive model), and 3) a third risk prediction model 230 that predicts the future risk of cancer within 5 years (e.g., 5 year, Lung-RADS 1-4B predictive model).

[0153] In certain embodiments in which a candidate subject is classified as Lung-RADS 4X, the risk expansion module 160 expands: 1) a first risk prediction model 230 that predicts the future risk of cancer within 1 year (e.g., 1 year, Lung-RADS 1-4X predictive model), 2) a second risk prediction model 230 that predicts the future risk of cancer within 3 years (e.g., 3 year, Lung-RADS 1-4X predictive model), and 3) a third risk prediction model 230 that predicts the future risk of cancer within 5 years (e.g., 5 year, Lung-RADS 1-4X predictive model).

[0154] 2A, risk prediction 240 can be used for patient enrichment 250. For example, subjects can be included in one or more patient cohorts enrolled in a clinical study. Methods for patient enrichment are described in more detail below.

[0155] Reference is now made to FIG. 2B, which shows an exemplary flow diagram for determining a subject's future risk of cancer for applications such as patient enrichment, according to a second embodiment. Here, FIG. 2B differs from FIG. 2A in that FIG. 2B does not include a subject analysis step (e.g., step 220 shown in FIG. 2A). Thus, FIG. 2B shows an embodiment in which subjects do not undergo screening. Instead, all subjects are candidate subjects who subsequently undergo future risk prediction. In various embodiments, subjects either have or do not have lung nodules. In various embodiments, subjects can be previously classified into any one of Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, Lung-RADS 4B, or Lung-RADS 4X. Thus, in such an embodiment, all subjects undergo future risk prediction, regardless of lung nodule staging and / or Lung-RADS score.

[0156] As shown in FIG. 2B, the candidate target image 255 undergoes feature extraction to obtain features 260. In various embodiments, the features 260 include one or both of the nodule-specific features and the non-nodule-specific features of the candidate target image 225. In various embodiments, the features 260 include only the non-nodule-specific features. In various embodiments, the features 260 include only the nodule-specific features. In various embodiments, the feature extraction module 140 extracts the nodule-specific features by implementing an image analysis algorithm, such as an image analysis algorithm that includes implementing a trained machine learning model. In various embodiments, the feature extraction module 140 extracts the nodule-specific features by implementing PyRadiomics. PyRadiomics is described in further detail in “Computational radiomics system to decode the radiographic phenotype.” Cancer Research; 77(21):e104-e107, the entire contents of which are incorporated herein by reference.

[0157] In various embodiments, the nodule-specific features of the candidate target images 255 are determined by a third party. For example, the nodule-specific features may be determined by a trained expert (e.g., a radiologist) analyzing the candidate target images 255. In various embodiments, the feature extraction module 145 extracts the nodule-specific features from a report generated by a third party. For example, the feature extraction module 145 extracts the nodule-specific features from a report generated by a trained expert (e.g., a radiologist) analyzing the target images 225.

[0158] The risk deployment module 260 applies one or more risk prediction models 265 to analyze the features 260 and generate a risk prediction 270. The risk prediction 270 may then be used for patient enrichment 275. In various embodiments, deploying the risk prediction model 265 to generate the risk prediction 270 as shown in FIG. 2B is the same process as deploying the risk prediction model 230 to generate the risk prediction 240 as shown in FIG. 2A.

[0159] Here, one or more risk prediction models 265 are trained with training images obtained from training individuals spanning the full range of possible subjects. For example, risk prediction models 265 are trained using training images captured from training individuals with or without lung nodules. As another example, risk prediction models 265 are trained using training images captured from training individuals with Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, Lung-RADS 4A, Lung-RADS 4B, or Lung-RADS 4X (e.g., Lung-RADS 1-4X risk prediction models).

[0160] In various embodiments, the risk expansion module 260 expands a plurality of future risk models 265 to predict a plurality of risk predictions 270. For example, the risk expansion module 260 expands a plurality of M-year, Lung-RADS 1-4X risk prediction models, where "M" refers to the time period over which cancer risk is being assessed. In various embodiments, "M" is at least 1 year, at least 2 years, at least 3 years, at least 4 years, at least 5 years, at least 6 years, at least 7 years, at least 8 years, at least 9 years, at least 10 years, at least 11 years, at least 12 years, at least 13 years, at least 14 years, at least 15 years, at least 16 years, at least 17 years, at least 18 years, at least 19 years, and / or at least 20 years. In certain embodiments, the risk expansion module 260 expands: 1) a 1-year, Lung-RADS 1-4X risk prediction model; 2) a 3-year, Lung-RADS 1-4X risk prediction model; and 3) a 5-year, Lung-RADS 1-4X risk prediction model, thereby generating future risk of cancer predictions for 1-year, 3-year, and 5-year time periods.

[0161] In various embodiments, once deployed, the risk prediction model analyzes the extracted image features and generates a prediction score that can indicate whether a subject is likely to develop cancer within a certain time period. For example, the risk prediction model can be a regression model (e.g., a logistic regression or linear regression model) that calculates a prediction score by combining a set of trained parameters with the values ​​of the extracted image features. As another example, the risk prediction model can be a neural network model that calculates a prediction score by combining a set of trained parameters associated with nodes and layers of the neural network with the values ​​of the extracted image features. As another example, the risk prediction model can be a random forest model that calculates a prediction score by combining a set of trained parameters associated with decision tree nodes with the values ​​of the extracted image features. As another example, the risk prediction model can be a gradient boosting machine model that calculates a prediction score by combining a set of trained parameters associated with decision tree nodes with the values ​​of the extracted image features.

[0162] In various embodiments, the risk prediction model compares the prediction score to one or more reference scores. In various embodiments, the one or more reference scores are threshold cutoff values. For example, the threshold cutoff value can be between 0 and 1, such as 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, or 0.9. In certain embodiments, the threshold is 0.1. In certain embodiments, the threshold is 0.3. Thus, if the prediction score exceeds the threshold reference score, the subject is classified into one category (e.g., more likely to develop cancer). If the prediction score is below the threshold reference score, the subject is classified into a different category (e.g., less likely to develop cancer).

[0163] In various embodiments, multiple reference threshold scores can be implemented to create multiple classification groups. For example, the first threshold is 0.1, and the second threshold is 0.3. Thus, if the prediction score is less than the first threshold, the subject is classified into a first category (e.g., low probability of developing cancer). If the prediction score is between the first and second thresholds, the subject is classified into a second category (e.g., low risk of developing cancer). If the prediction score is greater than the second threshold, the subject is classified into a third category (e.g., high risk of developing cancer).

[0164] As an example, the reference score corresponds to one or more training individuals. For example, the reference score can correspond to a training individual known to develop cancer within a time period. As another example, the reference score can correspond to a training individual known not to develop cancer within a time period. Thus, if the predicted score for the subject is not significantly different (e.g., p-value>0.05) compared to the reference score corresponding to the training individual known to develop cancer within a time period, the risk prediction model can classify the subject as likely to develop cancer within a time period. If the predicted score for the subject is not significantly different (e.g., p-value<0.05) compared to the reference score corresponding to the training individual known to develop cancer within a time period, the risk prediction model can classify the subject as not likely to develop cancer within a time period. If the predicted score for the subject is not significantly different (e.g., p-value>0.05) compared to the reference score corresponding to the training individual known not to develop cancer within a time period, the risk prediction model can classify the subject as not likely to develop cancer within a time period. If the prediction score for the subject is significantly different (e.g., p-value < 0.05) compared to the reference scores corresponding to training individuals who were known not to develop cancer within the time period, then the risk prediction model can classify the subject as likely to develop cancer within the time period.

[0165] In various embodiments, during training, a future cancer threshold risk is defined that distinguishes high-risk subjects from low-risk subjects. Then, only high-risk subjects are included in any given model. In various embodiments, each risk prediction model has a unique future cancer threshold that is used to distinguish subjects into high risk or low risk. Once those thresholds are defined for each risk prediction model, the risk prediction model is deployed for a subject to predict the future risk of cancer. If the risk exceeds the threshold defined in training, the subject can be classified as having a high future risk of cancer.

[0166] As shown in Figure 2B, risk prediction 270 can be used for patient enrichment 275. For example, subjects can be included in one or more patient cohorts enrolled in a clinical trial. Methods for patient enrichment are described in more detail below.

[0167] In various embodiments, the subject's risk prediction 270 can be displayed to a user, e.g., a clinician user. Thus, the clinician user can inform the subject of the predicted future risk of cancer for the subject. In various embodiments, additional / other information can be displayed to the user, e.g., a clinician user. For example, if the future risk of cancer prediction for the subject indicates that the subject is likely to develop cancer within a time period, information such as the features that contributed most to the future risk of cancer prediction can be displayed to the user, e.g., a clinician user. For example, a subject predicted to have a future risk of cancer can be largely attributed to the percentage of the subject's lungs that are occupied by centrilobular emphysema. Thus, the identity of the feature and / or the value of the feature (e.g., the percentage of the subject's lungs that are occupied by centrilobular emphysema) can be displayed to the user, e.g., a clinician user. In various embodiments, the top 1, top 2, top 3, top 4, top 5, top 6, top 7, top 8, top 9, or top 10 features that contributed most to the future risk of cancer prediction for the subject can be displayed to the user, e.g., a clinician user. Display of significant contributing features can provide context to clinician-users in understanding which features contributed to future risk of cancer prediction.

[0168] IV. Patient Enrichment Methods Generally, the future risk of cancer prediction from the risk prediction model is used for patient enrichment. For example, the future risk of cancer prediction provides insight into whether a subject is likely to develop cancer within a time period (e.g., within one year, within three years, or within five years). Thus, for subjects predicted to develop cancer within a certain time period, the subjects can be selected for inclusion in a patient cohort to be enrolled in a clinical trial. Given the insight provided by the future risk of cancer prediction, this allows for the enrollment of fewer subjects into the patient cohort of a clinical trial. Thus, fewer resources are required to conduct a clinical trial and follow up the subjects in the patient cohort. Furthermore, subjects not included in the patient cohort (e.g., subjects predicted not to develop cancer within a time period) can be used for other purposes (e.g., enrolled in other clinical trials).

[0169] In various embodiments, the subject is selected for inclusion in the patient cohort based on one or more of the multiple risk predictions generated for the patient.For example, the subject may have a first risk prediction indicating that the subject will not develop cancer within 1 year, will not develop cancer within 3 years, but will likely develop cancer within 5 years.Therefore, the subject is selected for inclusion in the patient cohort for enrolling in a cancer clinical trial that involves administration of a preventive therapeutic agent.

[0170] As another example, a subject may have a first risk prediction indicating that the subject is likely to develop cancer within one year, and therefore likely to develop cancer within three and five years. Thus, the subject is selected for inclusion in a patient cohort for enrollment in a cancer clinical trial involving aggressive cancer treatment (e.g., tumor resection and / or administration of a therapeutic agent).

[0171] In various embodiments, the patient enrichment process using risk predictions for multiple subjects generates an enriched cohort of patients who are more likely to develop cancer compared to a randomly generated patient cohort, hi various embodiments, the patient enrichment process generates an enriched cohort of patients who experience at least a 1.5-fold increase in cumulative cancer incidence compared to a randomly generated patient cohort. In various embodiments, the patient enrichment process generates an enriched cohort of patients that experience at least a 1.6-fold increase, at least a 1.7-fold increase, at least a 1.8-fold increase, at least a 1.9-fold increase, at least a 2-fold increase, at least a 3-fold increase, at least a 4-fold increase, at least a 5-fold increase, at least a 6-fold increase, at least a 7-fold increase, at least a 8-fold increase, at least a 9-fold increase, at least a 10-fold increase, at least an 11-fold increase, at least a 12-fold increase, at least a 13-fold increase, at least a 14-fold increase, at least a 15-fold increase, at least a 16-fold increase, at least a 17-fold increase, at least an 18-fold increase, at least a 19-fold increase, at least a 20-fold increase, at least a 21-fold increase, at least a 22-fold increase, at least a 23-fold increase, at least a 24-fold increase, at least a 25-fold increase, at least a 26-fold increase, at least a 27-fold increase, at least a 28-fold increase, at least a 29-fold increase, or at least a 30-fold increase in cumulative cancer incidence compared to a randomly generated patient cohort.

[0172] V. Training the risk prediction model Generally, the risk prediction model is configured to analyze features extracted from the image, such as non-nodule specific features and / or nodule specific features, and predict the future cancer risk of the subject based on the extracted features. In various embodiments, the risk prediction model is any one of a regression model (e.g., linear regression, logistic regression, or polynomial regression), a decision tree, a random forest, a gradient boosted machine learning model, a support vector machine, a naive Bayes model, a k-means cluster, or a neural network (e.g., a feed-forward network, a convolutional neural network (CNN), a deep neural network (DNN), an autoencoder neural network, a generative adversarial network, or a recurrent network (e.g., a long short-term memory network (LSTM), a bidirectional recurrent network, a deep bidirectional recurrent network), or any combination thereof. In certain embodiments, the risk prediction model is a logistic regression model. In certain embodiments, the risk prediction model is a random forest classifier. In certain embodiments, the risk prediction model is a gradient boosting model.

[0173] The risk prediction model can be trained using a machine learning implementation method, such as any one of a linear regression algorithm, a logistic regression algorithm, a decision tree algorithm, a support vector machine classification, a naive Bayes classification, a K-nearest neighbor classification, a random forest algorithm, a deep learning algorithm, a gradient boosting algorithm, and a dimensionality reduction technique such as manifold learning, principal component analysis, factor analysis, autoencoder regularization, and independent component analysis, or a combination thereof. In a particular embodiment, the machine learning implementation method is a logistic regression algorithm. In a particular embodiment, the machine learning implementation method is a random forest algorithm. In a particular embodiment, the machine learning implementation method is a gradient boosting algorithm such as XGboost. In various embodiments, the risk prediction model is trained using a supervised learning algorithm, an unsupervised learning algorithm, a semi-supervised learning algorithm (e.g., partially supervised), weakly supervised, transfer learning, multi-task learning, or any combination thereof.

[0174] In various embodiments, the risk prediction model has one or more parameters, such as hyperparameters or model parameters. The hyperparameters are generally established before training. Examples of hyperparameters include the learning rate, depth, or leaves of a decision tree, the number of hidden layers in a deep neural network, the number of clusters in a k-means cluster, the penalty in a regression model, and a regularization parameter associated with a cost function. The model parameters are generally adjusted during training. Examples of model parameters include weights associated with nodes in a layer of a neural network, support vectors in a support vector machine, node values ​​in a decision tree, and coefficients in a regression model. The model parameters of the risk prediction model are trained (e.g., adjusted) using training data to improve the predictive ability of the risk prediction model.

[0175] The risk training module 155 trains the risk prediction model using training data. The training data may be stored in and / or retrieved from the training data store 170. In various embodiments, the training data includes features extracted from training images obtained from training individuals (e.g., individuals known to develop or not develop cancer within a certain time period). In various embodiments, the training data may be obtained from a split of a dataset. For example, the dataset may undergo a 50:50 training:test dataset split. In some embodiments, the dataset may undergo a 60:40 training:test dataset split. In some embodiments, the dataset may undergo an 80:20 training:test dataset split.

[0176] In various embodiments, the training data used to train the imputation model includes a reference ground truth indicating whether a training individual developed cancer within the time period (hereinafter also referred to as "positive" or "+") or whether the training individual did not develop cancer within the time period (hereinafter also referred to as "negative" or "-"). In various embodiments, the reference ground truth in the training data is a binary value such as "1" or "0". For example, training individuals who developed cancer within the time period can be identified in the training data having a value of "1", while training individuals who did not develop cancer within the time period can be identified in the training data having a value of "0". In various embodiments, the risk training module 155 trains the risk prediction model using the training data to minimize a loss function so that the risk prediction model can better predict outcomes (e.g., future presence or absence of cancer within the time period) based on inputs (e.g., extracted features of the training images). In various embodiments, the loss function is constructed for either lasso (least absolute shrinkage and selection operator, LASSO) regression, ridge regression, or ElasticNet regression. In various embodiments, the risk prediction model is a random forest model and is trained to minimize one of the Gini impurity or entropy metrics for feature splitting, thereby enabling the risk prediction model to more accurately predict future cancer risk.

[0177] In various embodiments, the training data can be obtained and / or derived from a publicly available database. For example, the training data can be obtained and / or derived from the National Lung Screening Test (NLST). In some embodiments, the training data can be obtained and collected independently of a publicly available database, for example, by capturing images from multiple training individuals. Such training data can be a custom dataset.

[0178] In various embodiments, the risk prediction model is trained using a particular cohort of training individuals. In various embodiments, the risk prediction model is trained using a cohort of training individuals without pulmonary nodules. In various embodiments, the risk prediction model is trained using a cohort of training individuals with pulmonary nodules. In various embodiments, the risk prediction model is a Lung-RADS 1 predictive model trained using a cohort of training individuals previously classified into Lung-RADS 1. In various embodiments, the risk prediction model is a Lung-RADS 2 predictive model trained using a cohort of training individuals previously classified into any one of Lung-RADS 1-2. In various embodiments, the risk prediction model is a Lung-RADS 1-3 predictive model trained using a cohort of training individuals previously classified into any one of Lung-RADS 1-3. In various embodiments, the risk prediction model is a Lung-RADS 1-4A predictive model trained using a cohort of training individuals previously classified into any one of Lung-RADS 1-4A. In various embodiments, the risk prediction model is a Lung-RADS 1-4B predictive model that is trained using a cohort of training individuals previously classified into any one of Lung-RADS 1-4B. In various embodiments, the risk prediction model is a Lung-RADS 1-4X predictive model that is trained using a cohort of training individuals previously classified into any one of Lung-RADS 1-4X.

[0179] In various embodiments, for each of the Lung-RADS 1 predictive model, the Lung-RADS 1-2 predictive model, the Lung-RADS 1-3 predictive model, the Lung-RADS 1-4A predictive model, the Lung-RADS 1-4B predictive model, or the Lung-RADS 1-4X predictive model, the risk prediction model is trained using a training cohort of training individuals, where a majority (e.g., greater than 50%) of the training individuals in the training cohort were previously classified as Lung-RADS 1. In various embodiments, for each of the Lung-RADS 1 predictive model, the Lung-RADS 1-2 predictive model, the Lung-RADS 1-3 predictive model, the Lung-RADS 1-4A predictive model, the Lung-RADS 1-4B predictive model, or the Lung-RADS 1-4X predictive model, the risk prediction model is trained using a training cohort of training individuals, where a majority (e.g., greater than 50%) of the training individuals in the training cohort were previously classified as Lung-RADS 2. In various embodiments, for each of the Lung-RADS 1, Lung-RADS 1-2, Lung-RADS 1-3, Lung-RADS 1-4A, Lung-RADS 1-4B, or Lung-RADS 1-4X predictive models, the risk prediction model is trained using a training cohort of training individuals, where a majority (e.g., greater than 50%) of the training individuals in the training cohort were previously classified as Lung-RADS 1 or Lung-RADS 2. As shown in Table 1, approximately 90% of individuals are likely to fall into Lung-RADS 1 or Lung-RADS 2. Thus, such individuals can be used to train any of the risk prediction models.

[0180] In various embodiments, the risk prediction model is an "M" year predictive model trained to predict future risk of cancer within "M" years, such as within at least 1 year, at least 2 years, at least 3 years, at least 4 years, at least 5 years, at least 6 years, at least 7 years, at least 8 years, at least 9 years, at least 10 years, at least 11 years, at least 12 years, at least 13 years, at least 14 years, at least 15 years, at least 16 years, at least 17 years, at least 18 years, at least 19 years, or at least 20 years. Thus, the training data used to train the imputation model includes reference ground truth indicating that training individuals have developed cancer within "M" years.

[0181] In various embodiments, the risk prediction model is 1) trained using a particular cohort of training individuals and 2) trained to predict future risk of cancer in "M" years. As described above, the cohort of training individuals may be pre-classified into a range of Lung-RADS scores (e.g., Lung-RADS X through Y). Thus, the risk prediction model may be an M-year, Lung-RADS X through Y predictive model that is 1) trained using a cohort of training individuals classified as Lung-RADS X through Y and 2) trained to predict future risk of cancer in "M" years.

[0182] In various embodiments, the risk prediction model includes both non-nodule-specific features and nodule-specific features. Thus, when training the risk prediction model, the risk prediction model analyzes both non-nodule-specific features and nodule-specific features extracted from the training images and attempts to generate predictions that minimize a loss function. In general, the features of the risk prediction model have an importance value that reflects how much each feature influences the predictions generated by the risk prediction model. For example, a higher importance value of a feature indicates that the feature has a greater influence on the predictions generated by the risk prediction model compared to a different feature having a lower importance value.

[0183] In various embodiments, nodule-specific features of a risk prediction model have a greater impact on predicting future risk of cancer compared to non-nodule-specific features. For example, nodule-specific features of a risk prediction model have a higher feature importance value than non-nodule-specific features of a risk prediction model. In general, nodule-specific features have a higher importance value for risk prediction models trained to predict risk of cancer within a shorter time period (e.g., one year as opposed to three or five years). Furthermore, nodule-specific features have a higher importance value for risk prediction models trained using higher-risk lung cancer patients (e.g., patients classified as Lung-RADS 4A or Lung-RADS 4B).

[0184] In various embodiments, the feature with the highest importance value of the risk prediction model is a nodule-specific feature. In various embodiments, the top two features with highest importance values ​​of the risk prediction model are nodule-specific features. In various embodiments, the top three features with highest importance values ​​of the risk prediction model are nodule-specific features. In various embodiments, the top four features with highest importance values ​​of the risk prediction model are nodule-specific features. In various embodiments, the top 5, 6, 7, 8, 9, or 10 features with highest importance values ​​of the risk prediction model are nodule-specific features.

[0185] In various embodiments, more than 50% of the top three features with the highest feature importance values ​​are nodule-specific features. In various embodiments, two of the top three features with the highest feature importance values ​​are nodule-specific features. In various embodiments, three of the top five features with the highest feature importance values ​​are nodule-specific features. In various embodiments, four of the top five features with the highest feature importance values ​​are nodule-specific features. In various embodiments, more than 50% of the top five features with the highest feature importance values ​​are nodule-specific features. In various embodiments, more than 50% of the top ten features with the highest feature importance values ​​are nodule-specific features. In various embodiments, 6, 7, 8, 9, 10 of the top ten features with the highest feature importance values ​​are nodule-specific features.

[0186] In various embodiments, such a risk prediction model (e.g., a risk prediction model in which the nodule-specific features of the risk prediction model have a higher feature importance value than the non-nodule-specific features of the risk prediction model) is a one-year risk prediction model (e.g., predicts the risk of developing cancer within one year). In various embodiments, such a risk prediction model is a model trained with training images that include high-risk lung nodules. In various embodiments, such a risk prediction model can be trained with training images derived from individuals classified as Lung-RADS 4A. In various embodiments, such a risk prediction model can be trained with training images derived from individuals classified as Lung-RADS 4B. In various embodiments, such a risk prediction model can be trained with training images derived from individuals classified as Lung-RADS 4X. In various embodiments, such a risk prediction model can be trained with training images derived from individuals classified as Lung-RADS 4A, Lung-RADS 4B, Lung-RADS 4X, or any combination thereof. For example, such a risk prediction model can be any one of the Lung-RADS 1-4A, Lung-RADS 1-4B, or Lung-RADS 1-4X prediction models.

[0187] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the nodule-specific features of the risk prediction model have higher feature importance values ​​than the non-nodule-specific features of the risk prediction model) is a 1-year Lung-RADS 1-4B prediction model. For example, as shown in Table 3, in terms of feature importance for the 1-year Lung-RADS 1-4B prediction model, the top three features are nodule-specific features. Furthermore, in terms of feature importance for the 1-year Lung-RADS 1-4B prediction model, four of the top five features are nodule-specific features.

[0188] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the nodule-specific features of the risk prediction model have a higher feature importance value than the non-nodule-specific features of the risk prediction model) is a 3-year Lung-RADS 1-4B prediction model. For example, as shown in Table 3, in terms of feature importance for the 3-year Lung-RADS 1-4B prediction model, two of the top three features are nodule-specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 1-4B prediction model, three of the top five features are nodule-specific features. As another example, as shown in Table 19, in terms of feature importance for the 3-year Lung-RADS 1-4B prediction model, the top three features are nodule-specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 1-4B prediction model, four of the top five features are nodule-specific features.

[0189] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the nodule-specific features of the risk prediction model have higher feature importance values ​​than the non-nodule-specific features of the risk prediction model) is a 5-year, Lung-RADS 1-4B prediction model. For example, as shown in Table 3, two of the top three features in terms of feature importance for the 5-year, Lung-RADS 1-4B prediction model are nodule-specific features.

[0190] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the nodule-specific features of the risk prediction model have higher feature importance values ​​than the non-nodule-specific features of the risk prediction model) is a 1-year Lung-RADS 1-4A prediction model. For example, as shown in Table 6, in terms of feature importance for the 1-year Lung-RADS 1-4A prediction model, the top three features are nodule-specific features. Furthermore, in terms of feature importance for the 1-year Lung-RADS 1-4A prediction model, four of the top five features are nodule-specific features.

[0191] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the nodule-specific features of the risk prediction model have a higher feature importance value than the non-nodule-specific features of the risk prediction model) is a 3-year Lung-RADS 1-4A prediction model. For example, as shown in Table 6, in terms of feature importance for the 3-year Lung-RADS 1-4A prediction model, the top three features are nodule-specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 1-4A prediction model, four of the top five features are nodule-specific features. As another example, as shown in Table 20, in terms of feature importance for the 3-year Lung-RADS 1-4A prediction model, the top three features are nodule-specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 1-4A prediction model, four of the top five features are nodule-specific features.

[0192] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the nodule-specific features of the risk prediction model have higher feature importance values ​​than the non-nodule-specific features of the risk prediction model) is a 5-year, Lung-RADS 1-4A prediction model. For example, as shown in Table 6, two of the top three features in terms of feature importance for the 5-year, Lung-RADS 1-4A prediction model are nodule-specific features.

[0193] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the nodule-specific features of the risk prediction model have higher feature importance values ​​than the nodule-specific features of the risk prediction model) is a 1-year, Lung-RADS 1-3 prediction model. For example, as shown in Table 9, two of the top three features in terms of feature importance for the 1-year, Lung-RADS 1-3 prediction model are nodule-specific features.

[0194] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the nodule-specific features of the risk prediction model have higher feature importance values ​​than the non-nodule-specific features of the risk prediction model) is a 3-year Lung-RADS 2-4B prediction model. For example, as shown in Table 22, in terms of feature importance for the 3-year Lung-RADS 2-4B prediction model, the top three features are nodule-specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 2-4B prediction model, the top five features are nodule-specific features.

[0195] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the nodule-specific features of the risk prediction model have higher feature importance values ​​than the non-nodule-specific features of the risk prediction model) is a 3-year Lung-RADS 4A-4B prediction model. For example, as shown in Table 23, in terms of feature importance for the 3-year Lung-RADS 4A-4B prediction model, the top three features are nodule-specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 4A-4B prediction model, the top five features are nodule-specific features.

[0196] In various embodiments, non-nodal specific features of a risk prediction model have a greater impact on predicting future risk of cancer compared to nodal specific features. For example, non-nodal specific features of a risk prediction model have a higher feature importance value than nodal specific features of a risk prediction model. In general, non-nodal specific features have a higher importance value for risk prediction models trained to predict risk of cancer within a longer time period (e.g., 3 or 5 years as opposed to 1 year). Furthermore, non-nodal specific features have a higher importance value for risk prediction models trained using patients with a low risk of lung cancer (e.g., patients classified as Lung-RADS 2 or 3) or patients who do not yet have nodules (e.g., Lung-RADS 1).

[0197] In various embodiments, the feature with the highest importance value of the risk prediction model is a non-nodule specific feature. In various embodiments, the top two features with highest importance values ​​of the risk prediction model are non-nodule specific features. In various embodiments, the top three features with highest importance values ​​of the risk prediction model are non-nodule specific features. In various embodiments, the top four features with highest importance values ​​of the risk prediction model are non-nodule specific features. In various embodiments, the top 5, 6, 7, 8, 9, or 10 features with highest importance values ​​of the risk prediction model are non-nodule specific features.

[0198] In various embodiments, more than 50% of the top three features with the highest feature importance values ​​are non-nodule specific features. In various embodiments, two of the top three features with the highest feature importance values ​​are non-nodule specific features. In various embodiments, three of the top five features with the highest feature importance values ​​are non-nodule specific features. In various embodiments, four of the top five features with the highest feature importance values ​​are non-nodule specific features. In various embodiments, more than 50% of the top five features with the highest feature importance values ​​are non-nodule specific features. In various embodiments, more than 50% of the top ten features with the highest feature importance values ​​are non-nodule specific features. In various embodiments, 6, 7, 8, 9, 10 of the top ten features with the highest feature importance values ​​are non-nodule specific features.

[0199] In various embodiments, a risk prediction model that includes non-nodal specific features that have a greater influence on predicting future risk of cancer compared to nodal specific features is an M-year risk prediction model (e.g., predicts the risk of developing cancer within M years), where M is not equal to 1 year. In various embodiments, M is 2 years or more. For example, M can be 1.5 years, 2 years, 2.5 years, 3 years, 3.5 years, 4 years, 4.5 years, 5 years, 5.5 years, 6 years, 6.5 years, 7 years, 7.5 years, 8 years, 8.5 years, 9 years, 9.5 years, 10 years, 10.5 years, 11 years, 11.5 years, 12 years, 12.5 years, 13 years, 13.5 years, 14 years, 14.5 years, 15 years, 15.5 years, 16 years, 16.5 years, 17 years, 17.5 years, 18 years, 18.5 years, 19 years, 19.5 years, or 20 years. As a specific example, the risk prediction model is a 3-year risk prediction model that predicts the risk of developing cancer within 3 years. As another specific example, the risk prediction model is a 5-year risk prediction model that predicts the risk of developing cancer within 5 years.

[0200] In various embodiments, a risk prediction model including non-nodule specific features that have a greater impact on predicting future risk of cancer compared to nodule specific features is a risk prediction model trained with training images that do not include lung nodules. In various embodiments, such a risk prediction model can be trained with training images derived from individuals classified as Lung-RADS 1. In various embodiments, such a risk prediction model can be trained with training images derived from individuals classified as Lung-RADS 2. In various embodiments, such a risk prediction model can be trained with training images derived from individuals classified as Lung-RADS 3. In various embodiments, such a risk prediction model can be trained with training images derived from individuals classified as Lung-RADS 1, Lung-RADS 2, Lung-RADS 3, or any combination thereof. In various embodiments, such a risk prediction model is a Lung-RADS 1 prediction model, a Lung-RADS 1-2 prediction model, or a Lung-RADS 1-3 prediction model.

[0201] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have higher feature importance values ​​than the nodal specific features of the risk prediction model) is a 1-year Lung-RADS 1-4B prediction model. For example, as shown in Table 3, 6 of the top 10 features in terms of feature importance for the 1-year Lung-RADS 1-4B prediction model are non-nodal specific features.

[0202] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have higher feature importance values ​​than the nodal specific features of the risk prediction model) is a 3-year, Lung-RADS 1-4B prediction model. For example, as shown in Table 3, 6 of the top 10 features in terms of feature importance for the 3-year, Lung-RADS 1-4B prediction model are non-nodal specific features.

[0203] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have higher feature importance values ​​than the nodal specific features of the risk prediction model) is a 5-year Lung-RADS 1-4B prediction model. For example, as shown in Table 3, three of the top five features in terms of feature importance for the 5-year Lung-RADS 1-4B prediction model are non-nodal specific features. Furthermore, seven of the top ten features in terms of feature importance for the 5-year Lung-RADS 1-4B prediction model are non-nodal specific features.

[0204] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have higher feature importance values ​​than the nodal specific features of the risk prediction model) is a 5-year Lung-RADS 1-4A prediction model. For example, as shown in Table 6, three of the top five features in terms of feature importance for the 5-year Lung-RADS 1-4A prediction model are non-nodal specific features. Furthermore, six of the top ten features in terms of feature importance for the 5-year Lung-RADS 1-4A prediction model are non-nodal specific features.

[0205] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have higher feature importance values ​​than the nodal specific features of the risk prediction model) is a 1-year Lung-RADS 1-3 prediction model. For example, as shown in Table 9, three of the top five features in terms of feature importance for the 1-year Lung-RADS 1-3 prediction model are non-nodal specific features. Furthermore, seven of the top ten features in terms of feature importance for the 1-year Lung-RADS 1-3 prediction model are non-nodal specific features.

[0206] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have a higher feature importance value than the nodal specific features of the risk prediction model) is a 3-year Lung-RADS 1-3 prediction model. For example, as shown in Table 9, in terms of feature importance for the 3-year Lung-RADS 1-3 prediction model, two of the top three features are non-nodal specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 1-3 prediction model, four of the top five features are non-nodal specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 1-3 prediction model, eight of the top ten features are non-nodal specific features. For example, as shown in Table 21, in terms of feature importance for the 3-year Lung-RADS 1-3 prediction model, two of the top three features are non-nodal specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 1-3 predictive model, 3 of the top 5 features are non-nodal specific features.Furthermore, in terms of feature importance for the 3-year Lung-RADS 1-3 predictive model, 8 of the top 10 features are non-nodal specific features.

[0207] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have a higher feature importance value than the nodal specific features of the risk prediction model) is a 5-year Lung-RADS 1-3 prediction model. For example, as shown in Table 9, in terms of feature importance for the 5-year Lung-RADS 1-3 prediction model, two of the top three features are non-nodal specific features. Furthermore, in terms of feature importance for the 5-year Lung-RADS 1-3 prediction model, four of the top five features are non-nodal specific features. Furthermore, in terms of feature importance for the 5-year Lung-RADS 1-3 prediction model, nine of the top ten features are non-nodal specific features.

[0208] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have higher feature importance values ​​than the nodal specific features of the risk prediction model) is a 1-year Lung-RADS 1-2 prediction model. For example, as shown in Table 12, in terms of feature importance for the 1-year Lung-RADS 1-2 prediction model, four of the top five features are non-nodal specific features. Furthermore, in terms of feature importance for the 1-year Lung-RADS 1-2 prediction model, nine of the top ten features are non-nodal specific features.

[0209] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have higher feature importance values ​​than the nodal specific features of the risk prediction model) is a 3-year Lung-RADS 1-2 prediction model. For example, as shown in Table 12, in terms of feature importance for the 3-year Lung-RADS 1-2 prediction model, the top three features are non-nodal specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 1-2 prediction model, the top five features are non-nodal specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 1-2 prediction model, eight of the top ten features are non-nodal specific features.

[0210] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have higher feature importance values ​​than the nodal specific features of the risk prediction model) is a 5-year Lung-RADS 1-2 prediction model. For example, as shown in Table 12, in terms of feature importance for the 5-year Lung-RADS 1-2 prediction model, the top three features are non-nodal specific features. Furthermore, in terms of feature importance for the 5-year Lung-RADS 1-2 prediction model, the top five features are non-nodal specific features. Furthermore, in terms of feature importance for the 5-year Lung-RADS 1-2 prediction model, the top ten features are non-nodal specific features.

[0211] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have higher feature importance values ​​than the nodal specific features of the risk prediction model) is the 1-year Lung-RADS 1 prediction model. For example, as shown in Table 15, in terms of feature importance for the 1-year Lung-RADS 1 prediction model, the top three features are non-nodal specific features. Further, in terms of feature importance for the 1-year Lung-RADS 1 prediction model, the top five features are non-nodal specific features. Further, in terms of feature importance for the 1-year Lung-RADS 1 prediction model, the top ten features are non-nodal specific features.

[0212] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have higher feature importance values ​​than the nodal specific features of the risk prediction model) is a 3-year Lung-RADS 1 predictive model. For example, as shown in Table 15, in terms of feature importance for the 3-year Lung-RADS 1 predictive model, the top three features are non-nodal specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 1 predictive model, the top five features are non-nodal specific features. Furthermore, in terms of feature importance for the 3-year Lung-RADS 1 predictive model, the top ten features are non-nodal specific features.

[0213] In certain embodiments, such a risk prediction model (e.g., a risk prediction model in which the non-nodal specific features of the risk prediction model have higher feature importance values ​​than the nodal specific features of the risk prediction model) is a 5-year Lung-RADS 1 predictive model. For example, as shown in Table 15, in terms of feature importance for the 5-year Lung-RADS 1 predictive model, the top three features are non-nodal specific features. Furthermore, in terms of feature importance for the 5-year Lung-RADS 1 predictive model, the top five features are non-nodal specific features. Furthermore, in terms of feature importance for the 5-year Lung-RADS 1 predictive model, the top ten features are non-nodal specific features.

[0214] In various embodiments, the trained risk prediction model includes a set of trained parameters, and when the risk prediction model is deployed, the set of trained parameters is used to modify values ​​of non-nodule-specific and nodule-specific features of the image to generate a prediction of the subject's future risk of cancer. Thus, the set of trained parameters of the trained risk prediction model is set during the training phase. For example, the set of trained parameters is set such that non-nodule-specific features have a greater influence on the future risk prediction than nodule-specific features. As another example, the set of trained parameters is set such that nodule-specific features have a greater influence on the future risk prediction than non-nodule-specific features.

[0215] For example, if the risk prediction model is a neural network, one or more nodes of the neural network corresponding to non-nodule specific features are assigned a greater weight (e.g., parameter) than one or more nodes of the neural network corresponding to nodule specific features. As another example, the risk prediction model is a random forest model that weights non-nodule specific features more heavily than nodule specific features.

[0216] In various embodiments, the risk prediction model can also be trained to predict the future location of the cancer. In various embodiments, the training image is divided into different regions, and thus the training of the risk prediction model is performed according to the different regions. Exemplary different regions can include the upper third, middle third, and lower third of the lung by lung volume or separate lobes. It will be understood by those skilled in the art that the lung can be divided into any number of regions with any number of configurations. Enabling the risk prediction model to predict the future location of the cancer can guide the selection of interventions, such as regional diagnostic evaluations and therapeutic interventions using inhaled and bronchoscopically administered drugs and devices.

[0217] In various embodiments, the risk prediction models disclosed herein achieve a performance metric. Exemplary performance metrics include the area under the curve (AUC) of the receiver operating curve, positive predictive value, and / or negative predictive value. In various embodiments, the risk prediction models disclosed herein exhibit an AUC value of at least 0.5. In various embodiments, the risk prediction models disclosed herein exhibit an AUC value of at least 0.6. In various embodiments, the risk prediction models disclosed herein exhibit an AUC value of at least 0.7. In various embodiments, the risk prediction models disclosed herein exhibit an AUC value of at least 0.8. In various embodiments, the risk prediction models disclosed herein exhibit an AUC value of at least 0.9. In various embodiments, the risk prediction models disclosed herein exhibit an AUC value of at least 0.95. In various embodiments, the risk prediction models disclosed herein exhibit an AUC value of at least 0.99. In various embodiments, the risk prediction models disclosed herein provide a predictive probability of at least 0.51, at least 0.52, at least 0.53, at least 0.54, at least 0.55, at least 0.56, at least 0.57, at least 0.58, at least 0.59, at least 0.60, at least 0.61, at least 0.62, at least 0.63, at least 0.64, at least 0.65, at least 0.66, at least 0.67, at least 0.68, at least 0.69, at least 0.70, at least 0.71, at least 0.72, at least 0.73, at least 0.74, at least 0.75, at least 0.76, at least 0.77, at least 0.78, at least 0.79, at least 0.80, at least 0.81, at least 0.82, at least 0.83, at least 0.84, at least 0.85, at least 0.86, at least 0.87, at least 0.88, at least 0.89, at least 0.90, at least 0.91, at least 0.92, at least 0.93, at least 0.94, at least 0.95, at least 0.96, at least 0.97, at least 0.98, at least 0.99, at least 10 ... Also exhibits an AUC value of at least 0.74, at least 0.75, at least 0.76, at least 0.77, at least 0.78, at least 0.79, at least 0.80, at least 0.81, at least 0.82, at least 0.83, at least 0.84, at least 0.85, at least 0.86, at least 0.87, at least 0.88, at least 0.89, at least 0.90, at least 0.91, at least 0.92, at least 0.93, at least 0.94, at least 0.95, at least 0.96, at least 0.97, at least 0.98, at least 0.99.

[0218] VI. Exemplary Methods for Predicting Future Risk of Cancer 3 is an exemplary flow process for determining a subject's future risk of cancer for applications such as patient enrichment, according to one embodiment. Step 305 includes predicting future lung cancer risk for a plurality of subjects. Step 305 includes steps 310, 320, 330, and 340, which can be performed repeatedly for different subjects in the plurality of subjects.

[0219] An image is captured of the subject at step 310. In various embodiments, the image is a chest CT scan captured of the subject.

[0220] Features are extracted from the image at step 320. In various embodiments, the features include one or both of non-nodule specific features and nodule specific features.

[0221] Step 330 is an optional step that involves determining whether the subject is a candidate subject for future risk prediction based on one or more of the extracted features. In various embodiments, step 330 involves analyzing the nodule-specific features to determine whether the subject is a candidate subject. For example, step 330 can include analyzing the nodule-specific features to determine whether the subject does not have lung nodules or does not have lung cancer and is therefore eligible to undergo a future risk of cancer analysis.

[0222] Step 340 includes applying a risk prediction model to analyze the extracted features of the resulting images to predict future cancer risk. In various embodiments, the risk prediction model analyzes both non-nodule specific features and nodule specific features. In various embodiments, the non-nodule specific features have higher feature importance values ​​than the nodule specific features. Thus, the non-nodule specific features have a greater impact on future cancer risk prediction compared to the nodule specific features.

[0223] Step 350 includes enriching across subjects using the predicted future cancer risk, where subjects predicted to develop cancer within a period of time, as indicated by their predicted future cancer risk, can be included in one or more patient cohorts for enrollment in clinical trials. Collectively, this allows for a reduction in the number of individuals enrolled in clinical trials.

[0224] VII. Cancer The methods described herein include implementing a risk prediction model to predict future risk of cancer. In various embodiments, the cancer in the subject can include one or more of lymphoma, B-cell lymphoma, T-cell lymphoma, mycotic mycosis, Hodgkin's disease, myeloid leukemia, bladder cancer, brain tumor, cancer of the nervous system, head and neck cancer, head and neck squamous cell carcinoma, kidney cancer, lung cancer, neuroblastoma / glioblastoma, ovarian cancer, pancreatic cancer, prostate cancer, skin cancer, liver cancer, melanoma, squamous cell carcinoma of the mouth, throat, larynx, lung, colon cancer, cervical cancer, cervical carcinoma, breast cancer, and epithelial cancer, renal cancer, genitourinary cancer, lung cancer, esophageal cancer, gastric cancer, thyroid cancer, head and neck cancer, colon cancer, hematopoietic cancer, testicular cancer, colon and / or rectal cancer, uterine cancer, or prostate cancer. In some embodiments, the cancer in the subject may be a metastatic cancer, including any one of bladder cancer, breast cancer, colon cancer, kidney cancer, lung cancer, melanoma, ovarian cancer, pancreatic cancer, prostate cancer, rectal cancer, gastric cancer, thyroid cancer, or uterine cancer. In certain embodiments, the cancer is lung cancer. In certain embodiments, the cancer is a type of lung cancer, including any one of small cell lung cancer, non-small cell lung cancer, non-small cell carcinoma, adenocarcinoma, squamous cell carcinoma, large cell carcinoma, small cell carcinoma, combined small cell carcinoma, neuroendocrine tumor, pulmonary sarcoma, pulmonary lymphoma, and bronchial carcinoma.

[0225] In various embodiments, the risk prediction model described herein predicts the future risk of the presence of a cancer, such as lung cancer. In other words, the risk prediction model predicts the risk of the presence or absence of a cancer, such as lung cancer. In various embodiments, the risk prediction model described herein predicts the future risk of a subtype of lung cancer, including any one of small cell lung cancer, non-small cell lung cancer, non-small cell carcinoma, adenocarcinoma, squamous cell carcinoma, large cell carcinoma, small cell carcinoma, combined small cell carcinoma, neuroendocrine tumor, pulmonary sarcoma, pulmonary lymphoma, and bronchogenic carcinoma. In other words, the risk prediction model classifies a subject as likely to develop a particular subtype of lung cancer within a period of time (e.g., 1, 3, or 5 years). In certain embodiments, the risk prediction model predicts the future risk of non-small cell lung cancer or small cell lung cancer.

[0226] VIII. Intervention The embodiments described herein include implementing a risk prediction model to predict future risk of cancer. In various embodiments, an intervention is provided to the subject based on the future risk of cancer prediction. In various embodiments, the intervention can be any one of the following: application of a diagnosis, application of a preventive therapeutic agent, or a subsequent action. An exemplary subsequent action can include a subsequent examination of the subject to see if the subject develops cancer. The subsequent examination can include either a subsequent biopsy (e.g., a cancer biopsy or a lymph node biopsy) or a subsequent imaging scan (e.g., a CT scan, a PET scan, an MRI scan, an ultrasound imaging, or an X-ray imaging). In various embodiments, the subsequent examination of the subject can be at the next scheduled visit, or for a predetermined time (e.g., 1 month, 2 months, 3 months, 4 months, 5 months, 6 months, 7 months, 8 months, 9 months, 10 months, 11 months, 12 months, 18 months, or 24 months) after predicting the future risk of cancer. In various embodiments, the additional subsequent actions can include subsequent actions to treat the cancer present in the subject, such as tumor resection, bronchoscopic diagnosis, selection and / or administration of a therapeutic agent, selection / administration of a pharmaceutical composition, or any combination thereof.

[0227] In various embodiments, a therapeutic agent can be selected and / or administered to a subject based on a predicted future risk of cancer. The selected therapeutic agent is likely to delay or prevent the onset of cancer, such as lung cancer. Exemplary therapeutic agents include chemotherapy, energy therapy (e.g., external beam, microwave, radiofrequency ablation, brachytherapy, electroporation, cryoablation, photothermal ablation, laser therapy, photodynamic therapy, electrocautery, chemical embodied therapy, high intensity focused ultrasound, low intensity focused ultrasound), antigen-specific monoclonal antibodies, anti-inflammatory drugs, oncolytic virus therapy, or immunotherapy. In various embodiments, the selected therapeutic agent is an energy therapy, and the amount of energy applied (e.g., dose and duration) can be adjusted to achieve a desired therapeutic effect. In various embodiments, the therapeutic agent includes a small molecule or biologic, such as a cytokine, an antibody, a soluble cytokine receptor, an antisense oligonucleotide, an siRNA, and the like. Such biologic agents include mucins and derivatives of biologic agents, which can include, for example, fusion proteins, PEGylated derivatives, cholesterol-conjugated derivatives, and the like, known in the art. Also included are antagonists of cytokines and cytokine receptors, such as traps and monoclonal antagonists. Also included are biosimilar drugs of the active agents described herein or drugs that are bioequivalent to the active agents.

[0228] Drugs for treating lung cancer can include chemotherapy drugs such as docetaxel, cisplatin, carboplatin, gemcitabine, nab-paclitaxel, paclitaxel, pemetrexed, gefitinib, erlotinib, britinib (Alunbrig®), capmatinib (Tabrecta®), serpacatinib (Retevmo®), entrectinib (Rozlytrek®), lorlatinib (Lorbrena®), larotrectinib (Vitrakvi®), dacomitinib (Vizimpro®), and vinorelbine. Drugs for treating lung cancer can include antibody therapies such as durvalumab (Imfinzi®), nivolumab (Opdivo®), pembrolizumab (Keytruda®), atezolizumab (Tecentriq®), canakinumab, and ramucirumab.

[0229] In various embodiments, one or more of the described therapeutic agents can be combined as a combination therapy to treat a subject.

[0230] In various embodiments, pharmaceutical compositions can be selected and / or administered to a subject based on the subject's risk level for metastatic cancer, and the selected therapeutic agent is likely to show efficacy against the cancer. The pharmaceutical composition administered to an individual includes an active agent, such as a therapeutic agent described above. The active ingredient is present in a therapeutically effective amount, i.e., in an amount sufficient when administered to treat the disease or medical condition mediated thereby. The composition can also include various other agents for enhancing delivery and efficacy, to enhance delivery and stability of the active ingredient. Thus, for example, the composition can also include a pharmaceutically acceptable non-toxic carrier or diluent, which is defined as a vehicle commonly used to formulate pharmaceutical compositions for animal or human administration, depending on the desired formulation. The diluent is selected so as not to affect the biological activity of the combination. Examples of such diluents are distilled water, buffered water, saline, PBS, Ringer's solution, dextrose solution, and Hank's solution. In addition, the pharmaceutical composition or formulation can include other carriers, adjuvants, or non-toxic non-therapeutic, non-immunogenic stabilizers, excipients, and the like. The compositions may also include additional substances to approximate physiological conditions, such as pH adjusting and buffering agents, toxicity adjusting agents, wetting agents, detergents, etc. The compositions may also include any of a variety of stabilizing agents, such as antioxidants.

[0231] The pharmaceutical compositions or therapeutic agents described herein can be administered in a variety of different ways, examples include administering a composition containing a pharma- ceutical acceptable carrier via oral, intranasal, intramodular, intralesional, rectal, topical, intraperitoneal, intravenous, intramuscular, subcutaneous, transdermal, intrathecal, intrabronchial, transthoracic, or intracranial methods.

[0232] In various embodiments, a clinical response can be provided to a subject based on the predicted future risk of cancer generated for the subject by implementing a risk prediction model. In various embodiments, the clinical response can include providing counseling to modify the subject's behavior (e.g., counseling the patient about smoking cessation to reduce risk), initiating inhaled / topical, intravenous or enteral (oral) therapeutics that may delay / prevent malignant transformation, slow tumor growth or even prevent the spread (metastasis) of the disease, establishing an adaptive screening schedule for future risk similar to that performed in colonoscopy for polyps (e.g., individuals predicted to be at high risk for future lung cancer should undergo more frequent follow-up and imaging), or performing or scheduling to perform additional risk prediction tests to confirm the predicted future risk of lung cancer (e.g., individuals believed to be at high risk for lung cancer may undergo additional tests to confirm the risk or narrow down the type of cancer the individual is most at risk of developing). In various embodiments, additional risk prediction tests can include blood-based biomarkers (to look for non-specific inflammation, a known risk for lung cancer), metabolic / proteomic / gene expression / gene sequencing. The person can also have additional sampling of tissues (nasal epithelium, bronchial epithelium, etc.) to look at changes in gene expression in the airways.

[0233] IX. Computer Implementation The methods of the invention, including methods of implementing a risk prediction model for predicting future risk of cancer, are in some embodiments implemented on one or more computers.

[0234] For example, the construction and deployment of the risk prediction model can be implemented in hardware or software, or a combination of both. In one embodiment of the present invention, a machine-readable storage medium is provided, the medium comprising a data storage material encoded with machine-readable data that, when used with a machine programmed with instructions for using the data, can perform the training or deployment of the risk prediction model and / or display either the data set or the results (e.g., future risk of cancer prediction for a subject). The present invention can be embodied in a computer program executed on a programmable computer, the computer comprising a processor, a data storage system (including volatile and non-volatile memory and / or storage elements), a graphics adapter, a pointing device, a network adapter, at least one input device, and at least one output device. A display is coupled to the graphics adapter. The program code is applied to the input data to perform the functions described above and generate output information. The output information is applied to one or more output devices in a known manner. The computer can be, for example, a personal computer, a microcomputer, or a workstation of conventional design.

[0235] Each program may be implemented in a high-level procedural or object-oriented programming language to communicate with a computer system. However, if desired, the program may be implemented in assembly or machine language. In either case, the language may be a compiled or interpreted language. Each such computer program is preferably stored in a general-purpose or special-purpose programmable computer-readable storage medium or device (e.g., a ROM or magnetic diskette) to configure and operate the computer when the storage medium or device is read by the computer to perform the procedures described herein. The system may also be considered to be implemented as a computer-readable storage medium configured with a computer program, the storage medium so configured causing the computer to operate in a specific predefined manner to perform the functions described herein.

[0236] The signature patterns and their databases can be provided on a variety of media to facilitate their use. "Media" refers to a product containing the signature pattern information of the present invention. The database of the present invention can be recorded on a computer-readable medium, e.g., any medium that can be read and directly accessed by a computer. Such media include, but are not limited to, magnetic storage media such as floppy disks, hard disk storage media, and magnetic tapes, optical media such as CD-ROMs, electrical storage media such as RAM and ROM, and hybrids of these categories such as magnetic / optical storage media. Those skilled in the art can readily recognize how any of the currently known computer-readable media can be used to create a product containing a recording of the present database information. "Recorded" refers to a process for storing information on a computer-readable medium, using any such method as known in the art. Any convenient data storage structure can be selected based on the means used to access the stored information. A variety of data processor programs and formats can be used for storage, e.g., word processing text files, database formats, and the like.

[0237] In some embodiments, the methods of the present invention, including the method for predicting future risk of cancer by implementing a risk prediction model, are executed on one or more computers in a distributed computing system environment (e.g., in a cloud computing environment). In this description, "cloud computing" is defined as a model for enabling on-demand network access to a shared set of configurable computing resources. Cloud computing can be used to provide on-demand access to a shared set of configurable computing resources. The shared set of configurable computing resources can be rapidly provisioned via virtualization, released with low management effort or service provider interaction, and scaled accordingly. Cloud computing models can be configured with various characteristics, such as, for example, on-demand self-service, broad network access, resource pooling, rapid scalability, metered services, etc. Cloud computing models can also expose various service models, such as, for example, software as a service ("Software as a Service, SaaS"), platform as a service ("Platform as a Service, PaaS"), and infrastructure as a service ("Infrastructure as a Service, IaaS"). Cloud computing models may also be deployed using different deployment models such as private cloud, community cloud, public cloud, hybrid cloud, etc. In this description and claims, a "cloud computing environment" is an environment in which cloud computing is used.

[0238] Figure 4 illustrates an exemplary computer for implementing the entities illustrated in Figures 1A, 1B, 2A, 2B, and 3. Computer 400 includes at least one processor 402 coupled to a chipset 404. Chipset 404 includes a memory controller hub 420 and an input / output (I / O) controller hub 422. Memory 406 and graphics adapter 412 are coupled to memory controller hub 420, and display 418 is coupled to graphics adapter 412. Storage device 408, input device 414, and network adapter 416 are coupled to I / O controller hub 422. Other embodiments of computer 400 have different architectures.

[0239] The storage device 408 is a non-transitory computer-readable storage medium such as a hard drive, a compact disk read-only memory (CD-ROM), a DVD, or a solid-state memory device. The memory 406 holds instructions and data used by the processor 402. The input interface 414 is a touch screen interface, a mouse, a trackball, or other type of pointing device, a keyboard, or some combination thereof, and is used to input data into the computer 400. In some embodiments, the computer 400 may be configured to receive input (e.g., commands) from the input interface 414 via gestures from a user. The network adapter 416 couples the computer 400 to one or more computer networks.

[0240] The graphics adapter 412 displays images and other information on the display 418. In various embodiments, the display 418 is configured to allow a user (e.g., a radiologist, oncologist, pulmonologist) to input user selections on the display 418, for example, to initiate a patient risk prediction, order any additional tests or procedures, and / or set parameters of a risk prediction model. In one embodiment, the display 418 may include a touch screen. In various embodiments, the display 418 may show one or more future risks of cancer predictions for the subject. Thus, a user accessing the display 418 may inform the subject of the predicted future risk of cancer for the subject. In various embodiments, the display 418 may show information such as the features that contributed most significantly to the future risk of cancer prediction for the subject. For example, a subject predicted to have a future risk of cancer may be largely due to the percentage of the subject's lungs occupied by centrilobular emphysema. Thus, feature identification and / or feature value (e.g., percentage of the subject's lung occupied by centrilobular emphysema) can be shown to a user, e.g., a clinician user, on the display 418. In various embodiments, the top 1, top 2, top 3, top 4, top 5, top 6, top 7, top 8, top 9, or top 10 features that contributed most to the future risk of cancer prediction for the subject can be shown on the display 418. The display of the top contributing features can provide a context to a user, e.g., a clinician user, to understand which features contributed to the future risk of cancer prediction. The patient profile, CT images, generated risk assessment, and any other relevant information can be stored in memory so that the patient's information / results are accessible at any given time.

[0241] The computer 400 is adapted to execute computer program modules for providing the functionality described herein. As used herein, the term "module" refers to computer program logic used to provide a specified function. Thus, a module may be implemented in hardware, firmware, and / or software. In one embodiment, the program module is stored in the storage device 408, loaded into the memory 406, and executed by the processor 402.

[0242] The type of computer 400 used by the entities of Figure 1A or 1B can vary depending on the embodiment and the processing power required by the entities. For example, the cancer prediction system 130 can operate with a single computer 400 or multiple computers 400 that communicate with each other over a network, such as a server farm. The computer 400 can lack some of the components described above, such as the graphics adapter 412 and the display 418.

[0243] X. System Further disclosed herein is a system for implementing a risk prediction model for predicting future risk of cancer. In various embodiments, such a system may include at least the cancer prediction system 130 described above in FIG. 1A. In various embodiments, the cancer prediction system 130 is embodied as a computer system, such as the computer system having the exemplary computer 400 described in FIG. 4.

[0244] In various embodiments, the system includes an imaging device, such as imaging device 120 described above in FIG. 1A. In various embodiments, the system includes both a cancer prediction system 130 (e.g., a computer system) and an imaging device. In such embodiments, the cancer prediction system 130 can be communicatively coupled to the imaging device 120 to receive images (e.g., CT scans) captured from the subject. The computer system analyzes the images and implements a risk prediction model in silico to determine a future risk of lung cancer for the subject.

[0245] XI. Additional Embodiments Extensive research has been conducted to determine the earliest molecular changes that mark the transition from normal tissue repair to uncontrolled cell growth and cancer. Much of this work has focused on who is most resilient to or most susceptible to smoking-induced lung injury. These patterns of injury include distal small airway disease causing expiratory airflow obstruction, emphysematous destruction of the lung parenchyma, abnormal deposition of collagen and scar tissue (interstitial changes and fibrosis), and pulmonary vascular remodeling. It is increasingly clear that the occurrence of a mixture of some of these patterns of lung injury, all unique signs of susceptibility, increases the risk of developing lung cancer.

[0246] Observational studies have repeatedly demonstrated that smokers who develop expiratory airflow obstruction (a defining feature of COPD, i.e., Chronic Obstructive Pulmonary Disease) are at increased risk for future lung cancer. That risk is further increased in people with COPD who have emphysematous destruction of their distal air spaces. Further investigations have linked interstitial lung disease and pulmonary fibrosis (another form of smoking-related lung injury) to increased risk of lung cancer, again supporting the general contention that a condition characterized by chronic damage and repair may lead to malignancy. Finally, there are extrapulmonary manifestations of susceptibility to chronic tobacco smoke exposure, such as muscle wasting, that may identify increased risk of disease, possibly not only because of the direct link between muscle tissue and cancer, but also because loss of lean mass and skeletal muscle tissue may be a reflection of inflammatory overflow from already damaged lungs. All of these pulmonary and extrapulmonary manifestations of damage from chronic tobacco smoke exposure are detectable with computed tomographic (CT) imaging.

[0247] The present disclosure may be understood with reference to the following description and accompanying figures. The present disclosure relates to systems and methods that integrate expertise in medical image analysis and applied advanced machine learning techniques to metrics extracted from medical images obtained in the National Lung Screening Trial (NLST). Features extracted from these medical images are used to generate a risk model to predict future lung cancer. In particular, the risk model identifies patterns / mixtures of features that can be used to identify smokers at increased risk for future lung cancer.

[0248] The National Lung Screening Trial (NLST) was a randomized controlled trial to determine whether annual CT scans could reduce lung cancer deaths. The primary hypothesis driving the study was that CT imaging provides higher resolution in vivo data to detect cancer (or smaller size nodules) at an earlier stage that is more amenable to treatment and cure. The study recruited approximately 53,000 smokers, half of whom were randomized to receive annual CT scans and the other half to receive annual chest X-rays (CXRs). Smokers who received annual CT scans experienced an approximately 20% reduction in mortality due to lung cancer. These exciting results led to a change in health care where lung cancer screening CT is now a reimbursed part of preventive medicine.

[0249] The NLST CT scan and clinical data are currently freely available to the biomedical community, and baseline (T0), 1-year (T1), and year 2 (T2) CT scans from 15,000 individuals were acquired. According to an exemplary embodiment, image analysis algorithms were applied to all of these scans to extract measures of emphysema, interstitial changes, preserved lung tissue, and pectoral muscle size (area), with the latter muscle measurement being used as a proxy for body composition or fat-free mass. In one embodiment, the 15,000 T0 CT scans were split in half to construct a dedicated training set (n=7,500) and test set (n=7,500) of data. Using objective features extracted from the CT images, as well as subjective reports of visually identified nodules and their features (size, shape, etc.), a robust model was created to predict future lung cancer. The model strategy was based on a random forest approach to develop and optimize decision trees in the training data to predict the desired outcome. This approach was used rather than selecting a fixed threshold for a particular burden of disease (i.e., >10% emphysema or >10% interstitial change) because there is almost certainly some combination of emphysema, interstitial change, and sarcopenia that identifies an elevated risk of lung cancer. For example, a smoker whose lungs are 25% emphysema may have the same future risk of lung cancer as a smoker with 2% emphysema, 12% interstitial change, and reduced pectoral muscle area.

[0250] Random forest-based models were independently trained to predict 1-, 3-, and 5-year risk of future lung cancer in smokers enrolled in the NLST. These models were then modified to allow prediction of incident (new) cancers, rather than just identifying those with prevalent (already existing) cancer on T0 CT scans. This final step utilizes visual data describing the characteristics of any lung nodules in the CT images. While the exemplary embodiment shows and describes a random forest-based model, one skilled in the art will appreciate that other modeling approaches, such as logistic regression and XGBoost, can also be utilized. It will also be appreciated that while the risk prediction model is specifically shown and described as providing 1-, 3-, and 5-year risk predictions, the prediction terms can be varied as desired.

[0251] The lifespan of lung cancer most commonly begins with the development of a lung nodule. This nodule may be cancerous or may be a benign overgrowth of tissue that subsequently undergoes malignant transformation. The radiological community, in an attempt to recognize this process and standardize medical management, has developed a scoring system to stage nodules found on CT scans by their likelihood of becoming cancerous. This staging system, called the Lung CT Screening, Reporting Data System (LungRADS), is based on the size of the nodule, the rate of growth of the nodule, and the appearance of the nodule. In general, the larger the nodule, the more rapidly it is growing, or the more irregular its appearance, the more likely it is to become cancerous.

[0252] In one embodiment, the LungRADS scoring system was used to remove from the risk model those people most likely to have widespread cancer at the time of the T0 scan. To do this, people with the highest LungRADS scores (4A, 4B, and 4X) were excluded from this particular analysis (but were included in other models), but people with smaller, lower risk nodes (2 and 3) that may become cancer at some point in the future were kept. Models for 1-, 3-, and 5-year risk prediction were trained after removing all people with LungRADS 4 nodes, and then applied to the study subcohort of the NLST (n=7,500).

[0253] The combination of radiological features can still identify people with a 10+% absolute 3-year risk of developing lung cancer. Furthermore, about 1 / 3 of those people had no nodules on the baseline T0 scan, suggesting that these models are not just predicting which nodules may become cancerous, but which people will develop nodules and then be diagnosed with lung cancer. Using this approach, a series of risk prediction models were generated. These models differed by time to cancer (1, 3, and 5-year risk) and the nature of the nodules that were excluded. Thus, this submission includes all models across the LungRADS stages, including LungRADS 4A, 4B, and 4X.

[0254] Additional radiological features such as vascular and airway calcification, pulmonary vascular morphology, and bone mineral density may also represent unique metrics of disease susceptibility that can be used to predict future lung cancer. Nodule-specific features may also be integrated into these models. Nodule-specific features may include, for example, nodule location, proximity to emphysema and interstitial changes, rate of change of nodule characteristics (both absolute and relative to normalized standards).

[0255] In further embodiments, these image-based models may be integrated into more comprehensive biomarker panels for risk assessment and nodule identification. These panels may include, but are not limited to, proteomics, genetics, gene expression, cell-free circulating tumor DNA, etc., and it is anticipated that optimal model performance may result in combination and highly sensitive and specific markers.

[0256] These risk prediction models are premised on a mixture of objective and subjective features extracted from CT images. Extensive prior research suggests that objective but not subjective features on CT vary by CT scanner brand, model, and image reconstruction software. Additional geographic variation in disease burden among subjects may reflect real differences in exposure to harmful exposures (i.e., some people may be more or less susceptible to exposure damage and may effectively be more likely to develop emphysema or interstitial changes in lung tissue). For these reasons, absolute thresholds were not used to determine the presence and severity of CT metrics of disease. Instead, data from each of the study subcohorts were normalized by subtracting the mean and dividing by the standard deviation of each covariate. As additional cohorts are aggregated for model refinement, data normalization was performed using all existing data (cross-cohort) or by selecting a subset of data, including but not limited to those believed to best reflect the patient's unique biology, exposure history, ethnicity, or type of medical image being processed. Once normalized, the data were used in a random forest-based approach to modeling.

[0257] In one embodiment, the objective features include, for example, lung parenchymal features (e.g., densitometry measurements of lung parenchyma and measurements of interstitial changes of lung parenchyma) and muscle / chest wall body composition measurements. Densitometry measurements of lung parenchyma may include, for example, the percentage of the lung occupied by (i) low attenuation areas (LAA), defined as the area / volume with attenuation lower than -950 Hounsfield Units (HU), and (ii) high attenuation areas (HAA), defined as the area / volume of the lung with attenuation between -600 HU and -250 HU, and the ratio of LAA in the upper lung zone to LAA in the lower lung zone. Measurements of interstitial changes of lung parenchyma may further include regional histogram measurements of lung parenchyma, for example, the percentage of the lung occupied by normal tissue, centrilobular emphysema, centrilobular nodule, ground glass, honeycomb, linear scar, nodular, reticular, subpleural line, other emphysema, and / or cysts. Body composition measurements of muscle structure / chest wall may include, for example, pectoralis major cross-sectional area, pectoralis minor cross-sectional area, pectoralis major lean cross-sectional area, pectoralis minor lean cross-sectional area, left or right pectoralis major or pectoralis minor collective cross-sectional area, and subcutaneous fat cross-sectional area.

[0258] Risk prediction models trained using data from the NLST provide a multimodal risk assessment and enrichment approach. For example, consider two patients who developed cancer within 3 years of enrollment in the NLST. Both had Lung-RADS 3 nodules at baseline, both were predicted by the algorithm to develop cancer within 3 years, and both developed IA adenocarcinoma within 3 years. In fact, the predicted probability of cancer development for both patients is within 2% of each other (84.8% for patient 1 and 86.4% for patient 2). Patient 1 was a 65-year-old white female ever smoker with a 76 pack-year smoking history, a BMI of 23.5, relatively small pectoral muscle mass, and significant upper lobe predominant emphysema (31% emphysema based on regional histogram analysis). A 76 pack-year smoking history is equivalent to smoking one pack (20 cigarettes) of cigarettes daily for 76 years or two packs daily for 38 years. In contrast, patient 2 was a 59-year-old white male ever smoker with a 43 pack-year smoking history, a BMI of 29.8, relatively preserved pectoral muscle mass, and minimal emphysema (2.6% emphysema by volume based on local histogram analysis), but significantly more interstitial features (7.1% compared to 4%). In this second case, this last feature and the strong relationship between stromal features and cancer likely led the algorithm to predict the development of cancer within 3 years. The clinical and radiological differences between these two cases highlight the ability of risk prediction models to identify patients with unique and distinct phenotypes likely to develop cancer.

[0259] According to further embodiments, the model used to predict the individual's risk of future cancer can also be used to predict the location of that future cancer. In one embodiment, the data is divided into thirds to represent the upper, middle, and lower thirds of the lung by volume and by lobe. Model training and testing can be performed and regional (by volume and by lobe by third) risk of future lung cancer can be calculated. These additional data allow for targeted regional diagnostic evaluation and therapeutic intervention using inhaled and bronchoscopically administered drugs and devices. It will be understood by those skilled in the art that the lung can be divided into any number of regions with any number of configurations.

[0260] A system according to an exemplary embodiment of the present disclosure predicts future lung cancer risk based on CT images of a patient's lungs. The system may include a processor along with a user interface, a display, and a memory. The processor includes a risk prediction model, as described above, which may provide a patient-specific assessment of the likelihood of developing cancer in the future. The risk prediction model may be implemented by the processor, for example, as lines of code executed by the processor, as firmware executed by the processor, or as a function of the processor, which is an application specific integrated circuit (ASIC). Although the system is shown and described as including a computing system including a single processor, a user interface, a display, and a memory, it will also be understood by those skilled in the art that the system may be comprised of a network of computing systems, each of which includes one or more of the above components. In one example, the risk prediction model may be executed via a central processor in a network accessible via several different user stations. Alternatively, the risk prediction model may be executed via one or more processors.

[0261] Risk prediction models created to identify people at highest risk of developing future cancers enable clinical trials focused on cancer prevention and treatment of high-risk individuals, as well as enable clinical care by identifying which people should have the most aggressive follow-up and screening to detect cancer at the earliest stage. The patient-specific risk assessment may be output on the display of one or more computing stations and displayed to a user. The patient-specific risk assessment may be provided in any of a variety of configurations. The risk assessment may include a predicted level of risk and / or a time frame of risk (e.g., 1, 3, and 5-year risk). In one embodiment, the CT scans are color-coded by person and region to enable visualization of the modeling data. Areas with the highest risk are colored red. Moderate risk areas are yellow, and low risk areas are green. As a result, a person at low risk of future lung cancer may have green shaded in the whole lung image, while a person at high risk of future lung cancer may have a CT image with areas of red (corresponding to high risk), yellow (moderate risk), and green (low risk). For example, these color-coded maps may be presented on displays in radiology reading rooms for chest radiologists (to aid in clinical interpretation), as well as on displays of image-based workstations used by pulmonologists, oncologists, and proceduralists for targeted therapy. According to one embodiment, these maps are generated using three colors, as well as a continuous gradient of colors from red to green, based on the use case and visualization mode. However, those skilled in the art will appreciate that maps may be generated using any of a variety of colors and configurations, so long as the generated map indicates to the user the predicted risk of future lung cancer.

[0262] A user (e.g., radiologist, oncologist, pulmonologist) can input user selections on the user interface, for example, to initiate a patient risk prediction, order any additional tests or procedures, and / or set parameters for the risk prediction model. The user interface may include input devices, such as, for example, a keyboard and / or a mouse. In one embodiment, the user interface may include a touch interface enabled on the display. The patient profile, CT images, generated risk assessment, and any other associations may be stored in memory such that the patient's information / results may be accessible via the user at any given time.

[0263] Image-based risk prediction may include prediction of 1-, 3-, and 5-year risk of lung cancer. This set of models is used to develop preventive therapies for lung cancer by enabling clinical trials and ultimately clinical treatment. Clinical trials are more efficient and feasible because event rates (future lung cancer) will be above background, thereby allowing for smaller cohort sizes and potentially reducing the number of patients who need to be treated to reduce a single case of lung cancer. Following identification of an effective treatment for lung cancer prevention, this model may be used outside of clinical trials as a biomarker for clinical treatment to identify those who could benefit most from the treatment. Because all treatments carry some degree of risk, knowing who is most likely to develop lung cancer will impact risk / benefit decision analysis and reimbursement associated with the administration of treatment in individuals.

[0264] According to further embodiments, the risk prediction model may provide a basis for identifying different risks of lung cancer cell types, including adenocarcinoma and squamous cell carcinoma. These differences may impact treatment decisions and patient outcomes based on the cell type in which the cancer arises.

[0265] The risk prediction models described herein provide a patient-level probability of developing lung cancer. These models can be modified using regional metrics of lung susceptibility to predict the future location of lung cancer in that individual. As bronchoscopic procedures are increasingly utilized to diagnose, treat, and prevent lung cancer, knowledge of the lobes and regions in which lung cancer occurs can facilitate diagnostic biopsies of suspicious lesions or nodules and / or localized administration of treatment to prevent malignant transformation.

[0266] The risk prediction model described herein can also be used in treatment selection and medical decision making. There are competing risks of disease development and death in all people. For example, smokers can develop lung cancer, heart disease or cerebrovascular disease, all of which can be aggressive and life-threatening. The risk prediction model detailed in this document can be integrated with other existing clinically utilized risk models, such as the Framingham Heart Study Risk Score (to predict future heart disease), to provide a more holistic assessment of how to maintain a patient's health and inform patients of their most pressing health care needs.

[0267] The system's processor processes the patient's CT images to extract features related to parenchymal damage and body composition. Note that additional modeling can expand the list of features to include other data, including but not limited to vascular and airway calcification, pulmonary vascular morphology, bone mineral density, and nodule-specific features such as velocity of growth and proximity to emphysema and interstitial changes. The data is normalized per subject level using the distribution of CT features collected in a training set. This training set includes multiple cohorts representing different geographic regions, exposures, and tumor biology. A risk prediction model was created by using random forests to predict future lung cancer risk (per person and per lung region). In one embodiment, the risk assessment, which may include lung color coding based on regional risk of future cancer (red, yellow, and green to represent high, medium, and low risk of future cancer), may be displayed to the user on the display.

[0268] Additional embodiments are disclosed herein including a method including extracting features from a CT image of the patient's lungs, normalizing data including the extracted features from the CT image of the patient's lungs, and using a risk prediction model to generate a risk assessment including a prediction of future risk of lung cancer. In various embodiments, the risk assessment includes an image of the patient's lungs that is color coded based on local risk of future cancer. In various embodiments, the risk prediction model identifies one of a pattern and mixture of lung features that indicate an increased risk for lung cancer. In various embodiments, the extracted features include one of emphysema, interstitial changes, preserved lung problems, and pectoral muscle size.

[0269] Further disclosed herein is a system including a non-transitory computer readable storage medium storing an executable program, and a processor that executes the executable program to cause the processor to extract features from CT images of the patient's lungs, normalize data including the features extracted from the CT images of the patient's lungs, and generate a risk assessment including a prediction of future risk of lung cancer using a risk prediction model. EXAMPLES

[0270] The following are examples of specific embodiments for carrying out the present invention. The examples are provided for illustrative purposes only and are not intended to limit the scope of the present invention in any way. Efforts have been made to ensure accuracy with respect to the numerical values ​​used (e.g., amounts, temperatures, etc.), but some experimental error and deviation should be allowed for.

[0271] Example 1: Summary of risk prediction model construction and analysis Computed tomography (CT) images from the National Lung Cancer Screening Trial (NLST) cohort were analyzed to predict the likelihood of cancer in patients over different future horizons (e.g., within 1 year, within 3 years, or within 5 years). The predicted future risk of cancer was used to identify a population of patients enriched for the development of lung cancer.

[0272] The National Lung Screening Trial (NLST) was a randomized controlled trial to determine whether annual CT scans could reduce deaths from lung cancer. The primary hypothesis driving this study was that CT imaging provides higher resolution in vivo data to detect cancer (or smaller size nodules) at an earlier stage that is more amenable to treatment and cure. The study recruited approximately 53,000 smokers, half of whom were randomized to receive annual CT scans and the other half to receive annual chest x-rays (CXRs). These smokers who received annual CT scans experienced an approximately 20% reduction in mortality due to lung cancer. These exciting results led to a change in health care where lung cancer screening CT is now a reimbursed part of preventive medicine.

[0273] The NLST CT images and clinical data are currently freely available to the biomedical community, and CT scans were obtained. Overall demographics and details of all patients are shown in Table 2. Image analysis algorithms were applied to the CT images to extract features such as emphysema, interstitial changes, preserved lung tissue, and pectoral muscle size (area), the latter muscle measurement being used as a proxy for body composition or fat-free mass. The CT scans were split in half to construct a dedicated training set of data and a test set of data. Using objective features extracted from the CT images (e.g., non-nodule specific features), as well as subjective features of visually identified nodules and their properties (e.g., nodule specific features such as nodule size, shape, etc.), a robust model was created to predict future lung cancer.

[0274] Features (e.g., variables) were extracted from the CT images. Features included were selected based on previous experience and on likely biological relevance.

[0275] For the risk prediction models described in Examples 2-11 below, the following non-nodule specific features were used: 1) Lung parenchymal density measurements a. Percentage of lung occupied by: i.- Low attenuation areas (LAA), defined as areas / volumes with attenuation less than 950 Hounsfield units (HU) ii. High attenuation areas (HAA) defined as the area / volume of the lung with attenuation between -600HU and -250HU. b. Ratio of LAA in the upper lung zone to LAA in the lower lung zone (ratio LAA) 2) Muscle tissue / chest wall body composition measurements A cross-sectional area of ​​the pectoralis major muscle b. Pectoralis minor muscle cross-sectional area c.Leaned cross-sectional area of ​​pectoralis major muscle d. Pectoralis minor muscle lean cross-sectional area e. Subcutaneous fat cross-sectional area (axial or coronal) 3) Local histogram measurements of lung parenchyma a. Percentage of lung occupied by: i. Normal tissue ii. Centrilobular emphysema iii. Centrilobular nodules iv. Frosted glass v.Honeycomb lung vi. Linear marks vii. Nodules viii. Reticulated ix. Subpleural line x. Other emphysema xi. Cyst

[0276] For the risk prediction models described in Examples 2-6 below, the following nodule specific features include: 1) Attenuation 2) Margin Description 3) Diameter 4) Lung CT Screening, Reporting and Data System (Lung-RADS) score

[0277] For the risk prediction models described in Examples 7-11 below, the following radiomics features (e.g., nodule-specific features) were extracted from the original CT images, the wavelet-transformed CT images, and the Gaussian-transformed CT images: 1) Primary statistics 2) 3D shape-based features 3) 2D shape-based features 4) Gray-level co-occurrence matrix 5) Gray level run length matrix 6) Gray level size zone matrix 7) Adjacent graytone difference matrix 8) Gray level dependency matrix.

[0278] The features used specifically do not include clinical characteristics, which means that the clinical characteristics of the overall cohort are NLST characteristics. Clinical characteristics by lung cancer prediction category are shown below.

[0279] Risk prediction models are predicted for a mixture of objective (e.g., non-nodule specific) and subjective (e.g., nodule specific) features extracted from CT images. Objective but not subjective features on CT vary by CT scanner brand, model and image reconstruction software. Additional geographic variation in disease burden among subjects may reflect real differences in exposure to harmful exposures (i.e., some people may be more or less susceptible to exposure damage and effectively more likely to develop emphysema or interstitial changes in lung tissue). For these reasons, absolute thresholds were not used to determine the presence and severity of CT metrics of disease. Instead, data from each of the study subcohorts were normalized by subtracting the mean and dividing by the standard deviation of each covariate. As additional cohorts are aggregated for model refinement, data normalization is performed using all existing data (cross-cohort) or by selecting a subset of data, including but not limited to those believed to best reflect the patient's specific biology, exposure history, ethnicity, or type of medical image being processed. Once normalized, the data were used in a random forest-based approach to modeling.

[0280] Data preprocessing was first performed, which included normalizing the values ​​of continuous features (e.g., by centering (subtracting the mean) and scaling (splitting by standard deviation)). Normalization of the test set was performed using information from the training set only; i.e., the mean and standard deviation of the training set were used to normalize both the training and test sets. Further preprocessing was performed including downsampling of the majority class (cancer-free) to account for imbalanced data, i.e., the fact that cancer diagnoses are relatively rare. Alternative approaches such as SMOTE and ROSE were considered, but had similar performance to downsampling with higher computational requirements.

[0281] Multiple modeling approaches were evaluated to build risk prediction models, including logistic regression, XGBoost, and random forest. Random forest-based models were independently trained to predict 1-, 3-, and 5-year risk of future lung cancer in smokers enrolled in the NLST. The model strategy was based on a random forest approach to develop and optimize decision trees in the training data to predict the desired outcome. This approach was used rather than selecting a fixed threshold for a given burden of disease (i.e., >10% emphysema or >10% interstitial changes) because there will almost certainly be some combination of emphysema, interstitial changes, and sarcopenia that identifies an elevated risk of lung cancer. For example, a smoker with lungs where 25% of the lungs are emphysema may have the same future risk of lung cancer as a smoker with 2% emphysema, 12% interstitial changes, and reduced pectoral muscle area. These models were then modified to allow for the prediction of incident (new) cancers, rather than just identifying those with prevalent (already existing) cancers common on CT scans. This final step utilizes visual data to describe the characteristics of any pulmonary nodules in the CT images. Although the exemplary embodiments show and describe a random forest-based model and a gradient boosting model, one skilled in the art will appreciate that other modeling approaches, such as logistic regression, may also be utilized. Also, while the risk prediction model is specifically shown and described as providing 1, 3, and 5 year risk predictions, it will be appreciated that the prediction terms may be varied as desired.

[0282] To build the risk prediction model, the cohort was split 50 / 50 into training and test groups. The model was trained on the test group with three iterations of adjustment using 10-fold cross-validation. Dihortization of predicted probabilities was performed by maximizing the F-score within the cross-sampled training cohort. Results included cancer diagnosis at 1, 3, and 5 years. Performance measurements and visualizations included presentation of ROC curves and enrichments for all time points (as described in the examples below). Cumulative incidence and demographics of enriched and non-enriched cohorts based on 3-year cancer predictions are also predicted / illustrated in the examples below.

[0283] Models were constructed to predict future risk of cancer for either the entire cohort of patients or for subgroups of patients. Subgroups of patients were categorized based on baseline grading / categorization of nodules as Lung-RADS<4B, <4A, <3, and <2. Lung-RADS<4B includes patients categorized as Lung-RADS 1-4A. Lung-RADS<4A includes patients categorized as Lung_RADS 1-3. Lung-RADS<3 includes patients categorized as Lung_RADS 1-2. Lung-RADS<2 includes patients categorized as Lung_RADS 1. Table 1 summarizes the characteristics for the different Lung-RADS classifications. Demographic information and patient characteristics of the full cohort (including Lung-RADS classification) are shown in Table 2.

[0284] A separate risk prediction model was created for each subgroup of patients, i.e. the performance specified is not the performance of a cohort-wide model simply applied to the described subgroup, but rather the performance of a subgroup-specific model trained on patients in that subgroup.

[0285] As shown in the following examples, all results represent performance / findings based on models developed / trained on training cohorts and then applied to test cohorts.In the enrichment table, risk prediction models for each time range (1 year, 3 years, and 5 years) were trained separately because the most important features for 1 year cancer risk are not necessarily the same as those for 3 and 5 years cancer risk.This means that the cumulative incidence at 3 years is not just the cumulative incidence from the top row (1 year) plus the number of cases in the interval.

[0286] When interpreting cumulative incidence function plots, the raw probabilities on the y-axis are significantly affected by changes in risk setting due to censoring and death. These plots are included primarily to demonstrate the changes in their shape by subgroup. Of particular note is the fact that there is less of a steep rise in cancer diagnoses in the first year, except for those with larger / more nodules of concern.

[0287] When reviewing the receiver operating characteristic (ROC) curves and area under the curve (AUC) values, note that these are primarily included for their convention and references used in the literature: since the models were adjusted for the area under the precision-recall curve (e.g., referred to as "adjusted risk prediction models"), the area under the ROC is not significantly improved with the adjustment process.

[0288] Considering potential eligibility for enrollment, it is noted that individuals with 4A and 4B nodules require early evaluation based on Lung-RADS criteria, however, a significant proportion of both, and especially 4A, are not prevalent cancers and therefore may be able to enroll in future potential incident cancer studies.

[0289] Relative enrollment rates and percentages are CT screening enrollment rates for that Lung-RADS subgroup. Absolute enrollment rates are based on the overall cohort size.

[0290] Example 2: Predicting future risk of Lung-RADS 1-4B Using the method described in Example 1, three separate models were constructed using the complete patient cohort. Specifically, the first model is a 1-year Lung-RADS 1-4B risk prediction model, the second model is a 3-year Lung-RADS 1-4B risk prediction model, and the third model is a 5-year Lung-RADS 1-4B risk prediction model. For each risk prediction model, a training set (e.g., 50% of the cohort) was used to train the risk prediction model, and a test set (e.g., the other 50% of the cohort) was used to test the risk prediction model.

[0291] The top 10 significant features for each of the 1-, 3-, and 5-year risk prediction models are shown in Table 3. Notably, most of the significant features for the 1-year risk prediction models include nodule-specific features, whereas most of the significant features for the 3- and 5-year risk prediction models are objective features (e.g., non-nodule-specific features such as body and / or lung parenchymal features).

[0292] In particular, for the 1-year Lung-RADS 1-4B predictive model, the top three features in terms of feature importance are nodule-specific features. Furthermore, for the 1-year Lung-RADS 1-4B predictive model, four of the top five features are nodule-specific features. Furthermore, for the 1-year Lung-RADS 1-4B predictive model, six of the top ten features are non-nodule-specific features. For the 3-year Lung-RADS 1-4B predictive model, two of the top three features in terms of feature importance are nodule-specific features. Furthermore, for the 3-year Lung-RADS 1-4B predictive model, three of the top five features are nodule-specific features. Furthermore, for the 3-year Lung-RADS 1-4B predictive model, six of the top ten features are non-nodule-specific features. For the 5-year Lung-RADS 1-4B prediction model, two of the top three features in terms of feature importance are nodule-specific features. Additionally, for the 5-year Lung-RADS 1-4B prediction model, three of the top five features in terms of feature importance are non-nodule-specific features. Additionally, for the 5-year Lung-RADS 1-4B prediction model, seven of the top ten features in terms of feature importance are non-nodule-specific features.

[0293] FIG. 5A shows the performance of the risk prediction models to predict the likelihood of cancer within 1 year across patients' Lung-RADS 1-4B. The base random forest risk prediction model showed an AUC value of 0.92, whereas the adjusted random forest risk prediction model showed an AUC value of 0.90. FIG. 5B shows the performance of the risk prediction models to predict the likelihood of cancer within 3 years across patients' Lung-RADS 1-4B. The base random forest risk prediction model showed an AUC value of 0.79, whereas the adjusted random forest risk prediction model showed an AUC value of 0.79. Furthermore, Table 4 documents the characteristics of patients' Lung-RADS 1-4B according to the cancer prediction determined by the 3-year risk prediction models.

[0294] Figure 5C shows the performance of risk prediction models to predict the likelihood of cancer within 5 years across patients' Lung-RADS 1-4B. The base random forest risk prediction model showed an AUC value of 0.74, whereas the adjusted random forest risk prediction model showed an AUC value of 0.74. Taken together, the results in Figures 5A-5C show that different risk prediction models can be constructed and deployed to predict the likelihood of cancer in patients' Lung-RADS 1-4B across different future horizontal time points.

[0295] FIG. 5D shows the 3-year cumulative incidence function across patients' Lung-RADS 1-4B patients, where the risk prediction model predicted cancer for 450 patients and did not predict cancer for the other 6643 patients. If this includes patients' Lung-RADS 1-4B (including the most risky lung cancer nodules, e.g., Lung-RADS 4A / 4B), the cumulative incidence function for the 450 predicted cancer patients reflects a higher incidence in the early months (e.g., 0-6 months) compared to a lower incidence in subsequent years.

[0296] Table 5 shows the enrichment results of the entire patient cohort using future cancer predictions from the 1-, 3-, or 5-year risk prediction models. Specifically, Table 5 shows the enrichment results of the risk prediction models compared to the natural incidence of cancer in the original cohort (referred to as the "Null model" in Table 5). Application of the risk prediction models significantly improves the cumulative incidence (last column of Table 5).

[0297] Specifically, for the 1-year model, the natural incidence rate ("Null model") has a cumulative incidence of 1.79 (e.g., 1.79% of patients in the cohort will be diagnosed with cancer within 1 year). Application of a 1-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence of 19.72 (e.g., the 1-year risk prediction model predicts that 19.72% of patients in the cohort will be diagnosed with cancer within 1 year). Thus, application of the 1-year risk prediction model achieves an 11-fold increase in cumulative incidence.

[0298] For the 3-year model, the natural incidence rate ("Null Model") has a cumulative incidence of 4.2 (e.g., 4.2% of patients in the cohort will be diagnosed with cancer within 3 years). Application of a 3-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence of 26.44 (e.g., the 3-year risk prediction model predicts that 26.44% of patients in the cohort will be diagnosed with cancer within 3 years). Thus, application of the 3-year risk prediction model achieves a 6.3-fold increase in cumulative incidence.

[0299] For the 5-year model, the natural incidence rate ("Null Model") has a cumulative incidence of 5.58 (e.g., 5.58% of patients in the cohort will be diagnosed with cancer within 5 years). Application of a 5-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence of 28.43 (e.g., the 5-year risk prediction model predicts that 28.43% of patients in the cohort will be diagnosed with cancer within 5 years). Thus, application of the 5-year risk prediction model achieves a 5.1-fold increase in cumulative incidence.

[0300] Taken together, Table 5 shows that various risk prediction models can be implemented to enrich Lung-RADS 1-4B patients, thereby reducing the number of patients that need to be enrolled in clinical trials.

[0301] Example 3: Predicting future risk in patients with Lung-RADS 1-4A Using the method described in Example 1, three separate models were constructed using Lung-RADS 1-4A patients. Specifically, the first model is a 1-year Lung-RADS 1-4A risk prediction model, the second model is a 3-year Lung-RADS 1-4A risk prediction model, and the third model is a 5-year Lung-RADS 1-4A risk prediction model. For each risk prediction model, a training set (e.g., 50% of the cohort) was used to train the risk prediction model, and a test set (e.g., the other 50% of the cohort) was used to test the risk prediction model.

[0302] The top 10 significant features for each of the 1-, 3-, and 5-year risk prediction models are shown in Table 6. Notably, five of the top 10 features for the 1-year risk prediction model include nodule-specific features, whereas the majority of significant features for the 3- and 5-year risk prediction models are objective features (e.g., non-nodule-specific features such as body and / or lung parenchymal features).

[0303] In particular, for the 1-year Lung-RADS 1-4A prediction model, the top three features in terms of feature importance are nodule-specific features. Furthermore, for the 1-year Lung-RADS 1-4A prediction model, four of the top five features in terms of feature importance are nodule-specific features. Furthermore, for the 1-year Lung-RADS 1-4A prediction model, five of the top ten features in terms of feature importance are nodule-specific features.

[0304] For the 3-year Lung-RADS 1-4A prediction model, two of the top three features in terms of feature importance are nodule-specific features.Furthermore, for the 3-year Lung-RADS 1-4A prediction model, four of the top five features in terms of feature importance are nodule-specific features.Furthermore, for the 3-year Lung-RADS 1-4A prediction model, five of the top ten features in terms of feature importance are nodule-specific features.

[0305] For the 5-year Lung-RADS 1-4A prediction model, two of the top three features in terms of feature importance are nodule-specific features. Additionally, for the 5-year Lung-RADS 1-4A prediction model, three of the top five features in terms of feature importance are non-nodule-specific features. Additionally, for the 5-year Lung-RADS 1-4A prediction model, eight of the top ten features in terms of feature importance are non-nodule-specific features.

[0306] FIG. 6A shows the performance of the risk prediction models for predicting the likelihood of cancer within 1 year across Lung-RADS 1-4A patients. The base random forest risk prediction model showed an AUC value of 0.81, whereas the adjusted random forest risk prediction model showed an AUC value of 0.85. FIG. 6B shows the performance of the risk prediction models for predicting the likelihood of cancer within 3 years across Lung-RADS 1-4A patients. The base random forest risk prediction model showed an AUC value of 0.73, whereas the adjusted random forest risk prediction model showed an AUC value of 0.72. Furthermore, Table 7 documents the characteristics of the Lung-RADS 1-4A cohort of patients according to cancer prediction as determined by the 3-year risk prediction models.

[0307] FIG. 6C shows the performance of the risk prediction models for predicting the likelihood of cancer within 5 years across Lung-RADS 1-4A patients. The base random forest risk prediction model showed an AUC value of 0.73, whereas the adjusted random forest risk prediction model showed an AUC value of 0.72. Taken together, the results in FIG. 6A-FIG. 6C show that different risk prediction models can be constructed and deployed to predict the likelihood of cancer in Lung-RADS 1-4A patient cohorts across different future horizontal time points. In other words, even if the patients at the highest risk (e.g., Lung-RADS 4B) are excluded, the risk prediction models can accurately predict the likelihood of future risk of cancer in lower risk patients (e.g., Lung-RADS 1-4A).

[0308] Figure 6D shows the 3-year cumulative incidence function across Lung-RADS 1-4A patients, where the risk prediction model predicted cancer for 392 patients and did not predict cancer for the other 6531 patients. Given that the patient cohort excluded Lung-RADS 4B patients, the cumulative incidence function of the 392 predicted cancer patients shown in Figure 6D reflects a lower incidence (e.g., less prevalent cancer or less cancer present at the time of the T0 CT scan) in early months (e.g., between 0 and 6 months) compared to Example 2 (shown in Figure 5D).

[0309] Table 8 shows the enrichment results of the Lung-RADS 1-4A patient cohort using future cancer predictions from the 1-, 3-, or 5-year risk prediction models. Specifically, Table 8 shows the enrichment results of the risk prediction models compared to the natural incidence of cancer in the original cohort (referred to as the "Null model" in Table 8). Application of the risk prediction models significantly improves the cumulative incidence (last column of Table 8).

[0310] Specifically, for the 1-year model, the natural incidence rate ("Null model") has a cumulative incidence of 0.98 (e.g., 0.98% of patients in the cohort will be diagnosed with cancer within 1 year). Application of a 1-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence of 7.48 (e.g., the 1-year risk prediction model predicts that 7.48% of patients in the cohort will be diagnosed with cancer within 1 year). Thus, application of the 1-year risk prediction model achieves a 7.6-fold increase in cumulative incidence.

[0311] For the 3-year model, the natural incidence rate ("Null Model") has a cumulative incidence of 3.28 (e.g., 3.28% of patients in the cohort will be diagnosed with cancer within 3 years). Application of a 3-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence of 16.84 (e.g., the 3-year risk prediction model predicts that 26.44% of patients in the cohort will be diagnosed with cancer within 3 years). Thus, application of the 3-year risk prediction model achieves a 5.1-fold increase in cumulative incidence.

[0312] For the 5-year model, the natural incidence rate ("Null Model") has a cumulative incidence of 4.65 (e.g., 4.65% of patients in the cohort will be diagnosed with cancer within 5 years). Application of a 5-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence of 17.97 (e.g., the 5-year risk prediction model predicts that 17.97% of patients in the cohort will be diagnosed with cancer within 5 years). Thus, application of the 5-year risk prediction model achieves a 3.9-fold increase in cumulative incidence.

[0313] Taken together, Table 8 shows that various future risk prediction models can be implemented to enrich patients in the Lung-RADS 1-4A cohort, thereby reducing the number of patients that need to be enrolled in clinical trials.

[0314] Example 4: Predicting future risk in patients with Lung-RADS 1-3 Using the method described in Example 1, three separate models were constructed using Lung-RADS 1-3 patients. Specifically, the first model was a 1-year Lung-RADS 1-3 risk prediction model, the second model was a 3-year Lung-RADS 1-3 risk prediction model, and the third model was a 5-year Lung-RADS 1-3 risk prediction model. For each risk prediction model, a training set (e.g., 50% of the cohort) was used to train the risk prediction model, and a test set (e.g., the other 50% of the cohort) was used to test the risk prediction model.

[0315] The top 10 significant features for each of the 1-, 3-, and 5-year risk prediction models are shown in Table 9. Notably, most of the significant features for the 1-year risk prediction models include nodule-specific features, whereas most of the significant features for the 3- and 5-year risk prediction models are objective features (e.g., non-nodule-specific features such as body and / or lung parenchymal features).

[0316] In particular, for the 1-year Lung-RADS 1-3 predictive model, two of the top three features in terms of feature importance are nodule-specific features. Additionally, for the 1-year Lung-RADS 1-3 predictive model, three of the top five features in terms of feature importance are non-nodule-specific features. Additionally, for the 1-year Lung-RADS 1-3 predictive model, seven of the top ten features in terms of feature importance are non-nodule-specific features.

[0317] For the 3-year Lung-RADS 1-3 predictive model, two of the top three features in terms of feature importance are non-nodal specific features.Furthermore, for the 3-year Lung-RADS 1-3 predictive model, four of the top five features in terms of feature importance are non-nodal specific features.Furthermore, for the 3-year Lung-RADS 1-3 predictive model, eight of the top ten features in terms of feature importance are non-nodal specific features.

[0318] For the 5-year Lung-RADS 1-3 predictive model, two of the top three features in terms of feature importance are non-nodal specific features.Furthermore, for the 5-year Lung-RADS 1-3 predictive model, four of the top five features in terms of feature importance are non-nodal specific features.Furthermore, for the 5-year Lung-RADS 1-3 predictive model, nine of the top ten features in terms of feature importance are non-nodal specific features.

[0319] FIG. 7A shows the performance of the risk prediction models for predicting the likelihood of cancer within 1 year across Lung-RADS 1-3 patients. The base random forest risk prediction model showed an AUC value of 0.82, whereas the adjusted random forest risk prediction model showed an AUC value of 0.82. FIG. 7B shows the performance of the risk prediction models for predicting the likelihood of cancer within 3 years across Lung-RADS 1-3 patients. The base random forest risk prediction model showed an AUC value of 0.72, whereas the adjusted random forest risk prediction model showed an AUC value of 0.73. Additionally, Table 10 documents the characteristics of the Lung-RADS 1-3 patient cohort of patients according to cancer prediction as determined by the 3-year risk prediction models.

[0320] FIG. 7C shows the performance of the risk prediction models for predicting the likelihood of cancer within 5 years across Lung-RADS 1-3 patients. The base random forest risk prediction model showed an AUC value of 0.67, whereas the adjusted random forest risk prediction model showed an AUC value of 0.68. Taken together, the results in FIG. 7A-7C show that different risk prediction models can be constructed and deployed to predict the likelihood of cancer in Lung-RADS 1-3 patient cohorts across different future horizontal time points. In other words, even if the patients at the highest risk (e.g., Lung-RADS 4A / 4B) are excluded, the risk prediction models can accurately predict the likelihood of future risk of cancer in lower risk patients (e.g., Lung-RADS 1-3).

[0321] FIG. 7D shows the 3-year cumulative incidence function across Lung-RADS 1-3 patients, where the risk prediction model predicted cancer for 324 patients and did not predict cancer for the other 6085 patients. Considering that the patient cohort excluded both Lung-RADS 4A and 4B patients, the exclusion of 4A and 4B patients manifests as a lower incidence of cancer in the 324 predicted cancer patients in early months (e.g., months 0-6) compared to the predicted cancer patients in Example 2 shown in FIG. 5D. Specifically, the cumulative incidence at the 1-year mark for the Lung-RADS 1-3 patient cohort is less than 0.1 (shown in FIG. 7D), while the cumulative incidence at the 1-year mark for the complete patient cohort (e.g., Lung-RADS 1-4B) is approximately 0.2.

[0322] Table 11 shows the enrichment results of the Lung-RADS 1-3 patient cohort using future cancer predictions from the 1-, 3-, or 5-year risk prediction models. Specifically, Table 11 shows the enrichment results of the risk prediction models compared to the natural incidence of cancer in the original cohort (referred to as the "Null model" in Table 11). Application of the risk prediction models significantly improves the cumulative incidence (last column of Table 11).

[0323] Specifically, for the 1-year model, the natural incidence rate ("Null model") has a cumulative incidence of 0.70 (e.g., 0.70% of patients in the cohort will be diagnosed with cancer within 1 year). Application of the 1-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence of 6.03 (e.g., the 1-year risk prediction model predicts that 6.03% of patients in the cohort will be diagnosed with cancer within 1 year). Thus, application of the 1-year risk prediction model achieves an 8.6-fold increase in cumulative incidence.

[0324] For the 3-year model, the natural incidence rate ("Null Model") has a cumulative incidence rate of 2.73 (e.g., 2.73% of patients in the cohort will be diagnosed with cancer within 3 years). Application of a 3-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence rate of 15.43 (e.g., the 3-year risk prediction model predicts that 15.43% of patients in the cohort will be diagnosed with cancer within 3 years). Thus, application of the 3-year risk prediction model achieves a 5.7-fold increase in cumulative incidence rate.

[0325] For the 5-year model, the natural incidence rate ("Null Model") has a cumulative incidence rate of 4.06 (e.g., 4.06% of patients in the cohort will be diagnosed with cancer within 5 years). Application of a 5-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence rate of 14.88 (e.g., the 5-year risk prediction model predicts that 14.88% of patients in the cohort will be diagnosed with cancer within 5 years). Thus, application of the 5-year risk prediction model achieves a 3.7-fold increase in cumulative incidence rate.

[0326] Taken together, Table 11 shows that various risk prediction models can be implemented to enrich patients in the Lung-RADS 1-3 cohort, thereby reducing the number of patients that need to be enrolled in clinical trials.

[0327] Example 5: Predicting future risk in patients with Lung-RADS 1-2 Using the method described in Example 1, three separate models were constructed using Lung-RADS 1-2 patients. Specifically, the first model was a 1-year Lung-RADS 1-2 risk prediction model, the second model was a 3-year Lung-RADS 1-2 risk prediction model, and the third model was a 5-year Lung-RADS 1-2 risk prediction model. For each risk prediction model, a training set (e.g., 50% of the cohort) was used to train the risk prediction model, and a test set (e.g., the other 50% of the cohort) was used to test the risk prediction model.

[0328] The top 10 significant features for each of the 1-, 3-, and 5-year risk prediction models are shown in Table 12. Notably, most of the significant features for the 1-year risk prediction models include nodule-specific features, whereas most of the significant features for the 3- and 5-year risk prediction models are objective features (e.g., non-nodule-specific features such as body and / or lung parenchymal features).

[0329] In particular, two of the top three features in terms of feature importance for the 1-year Lung-RADS 1-2 predictive model are nodule-specific features. Additionally, three of the top five features in terms of feature importance for the 1-year Lung-RADS 1-2 predictive model are non-nodule-specific features. Additionally, eight of the top ten features in terms of feature importance for the 1-year Lung-RADS 1-2 predictive model are non-nodule-specific features.

[0330] For the 3-year Lung-RADS 1-2 predictive model, the top three features in terms of feature importance are non-nodal specific features. Additionally, the top five features in terms of feature importance for the 3-year Lung-RADS 1-2 predictive model are non-nodal specific features. Additionally, eight of the top ten features in terms of feature importance for the 3-year Lung-RADS 1-2 predictive model are non-nodal specific features.

[0331] For the 5-year Lung-RADS 1-2 prediction model, the top 3 features in terms of feature importance are non-nodal specific features. Additionally, the top 5 features in terms of feature importance for the 5-year Lung-RADS 1-2 prediction model are non-nodal specific features. Additionally, the top 10 features in terms of feature importance for the 5-year Lung-RADS 1-2 prediction model are non-nodal specific features.

[0332] FIG. 8A shows the performance of the risk prediction models for predicting the likelihood of cancer within 1 year across Lung-RADS 1-2 patients. The base random forest risk prediction model showed an AUC value of 0.73, whereas the adjusted random forest risk prediction model showed an AUC value of 0.72. FIG. 8B shows the performance of the risk prediction models for predicting the likelihood of cancer within 3 years across Lung-RADS 1-2 patients. The base random forest risk prediction model showed an AUC value of 0.66, whereas the adjusted random forest risk prediction model showed an AUC value of 0.65. Additionally, Table 13 documents the characteristics of the Lung-RADS 1-2 patient cohort according to cancer prediction as determined by the 3-year risk prediction models.

[0333] Figure 8C shows the performance of the risk prediction models for predicting the likelihood of cancer within 5 years across Lung-RADS 1-2 patients. The base random forest risk prediction model showed an AUC value of 0.65, whereas the adjusted random forest risk prediction model showed an AUC value of 0.65. Taken together, the results of Figures 8A-8C show that different risk prediction models can be constructed and deployed to predict the likelihood of cancer in Lung-RADS 1-2 patient cohorts across different future horizontal time points. In other words, even if patients at higher risk (e.g., Lung-RADS 3 / 4A / 4B) are excluded, the risk prediction models can accurately predict the likelihood of future risk of cancer in low-risk patients (e.g., Lung-RADS 1-2). Notably, as shown in Table 1, the prevalence of Lung-RADS 1-2 individuals is approximately 90% of the population. Therefore, a risk prediction model that can predict the likelihood of cancer in Lung-RADS 1-2 patient cohorts would be highly valuable because it can be applied to a large portion of the patient population.

[0334] FIG. 8D shows the 3-year cumulative incidence function across Lung-RADS 1-2 patients. Here, the risk prediction model predicted cancer for 294 patients and did not predict cancer for the other 5314 patients. Considering that the patient cohort excluded Lung-RADS 3, 4A, and 4B patients, the exclusion of Lung-RADS 3, 4A, and 4B patients manifests as a lower incidence of cancer in the 294 predicted cancer patients in early months (e.g., months 0-6) compared to Example 2 shown in FIG. 5D. Specifically, the cumulative incidence at the 1-year mark for the Lung-RADS 1-2 patient cohort is less than 0.025 (shown in FIG. 8D), while the cumulative incidence at the 1-year mark for the complete patient cohort (e.g., Lung-RADS 1-4B) is approximately 0.2.

[0335] Table 14 shows the enrichment results of the Lung-RADS 1-2 patient cohort using future cancer predictions from the 1-, 3-, or 5-year risk prediction models. Specifically, Table 14 shows the enrichment results of the risk prediction models compared to the natural incidence of cancer in the original cohort (referred to as the "Null Model" in Table 14). Application of the risk prediction models significantly improves the cumulative incidence (last column of Table 14).

[0336] Specifically, for the 1-year model, the natural incidence rate ("Null model") has a cumulative incidence of 0.43 (e.g., 0.43% of patients in the cohort will be diagnosed with cancer within 1 year). Application of a 1-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence of 2.37 (e.g., the 1-year risk prediction model predicts that 2.37% of patients in the cohort will be diagnosed with cancer within 1 year). Thus, application of the 1-year risk prediction model achieves a 5.5-fold increase in cumulative incidence.

[0337] For the 3-year model, the natural incidence rate ("Null Model") has a cumulative incidence of 2.37 (e.g., 2.37% of patients in the cohort will be diagnosed with cancer within 3 years). Application of a 3-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence of 7.14 (e.g., the 3-year risk prediction model predicts that 7.14% of patients in the cohort will be diagnosed with cancer within 3 years). Thus, application of the 3-year risk prediction model achieves a 3-fold increase in cumulative incidence.

[0338] For the 5-year model, the natural incidence rate ("Null Model") has a cumulative incidence of 3.67 (e.g., 3.67% of patients in the cohort will be diagnosed with cancer within 5 years). Application of a 5-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence of 7.06 (e.g., the 5-year risk prediction model predicts that 7.06% of patients in the cohort will be diagnosed with cancer within 5 years). Thus, application of the 5-year risk prediction model achieves a 1.9-fold increase in cumulative incidence.

[0339] Taken together, Table 14 shows that various risk prediction models can be implemented to enrich patients in the Lung-RADS 1-2 cohort, thereby reducing the number of patients that need to be enrolled in clinical trials.

[0340] Example 6: Predicting future risk in Lung-RADS 1 patients Using the method described in Example 1, three separate models were constructed using Lung-RADS 1 patients. Specifically, the first model is a 1-year, Lung-RADS 1 risk prediction model, the second model is a 3-year, Lung-RADS 1 risk prediction model, and the third model is a 5-year risk, Lung-RADS 1 prediction model. For each risk prediction model, a training set (e.g., 50% of the cohort) was used to train the risk prediction model, and a test set (e.g., the other 50% of the cohort) was used to test the risk prediction model.

[0341] The top 10 significant features for each of the 1-, 3-, and 5-year risk prediction models are shown in Table 15. Notably, most of the significant features for the 1-year risk prediction models include nodule-specific features, whereas most of the significant features for the 3- and 5-year risk prediction models are objective features (e.g., non-nodule-specific features such as body and / or lung parenchymal features).

[0342] In particular, for the 1-year Lung-RADS 1 predictive model, the top three features in terms of feature importance are non-nodal specific features. Furthermore, the top five features in terms of feature importance for the 1-year Lung-RADS 1 predictive model are non-nodal specific features.

[0343] For the 3-year Lung-RADS 1 predictive model, the top 3 features in terms of feature importance are non-nodal specific features. Additionally, the top 5 features in terms of feature importance for the 3-year Lung-RADS 1 predictive model are non-nodal specific features. Additionally, the top 10 features in terms of feature importance for the 3-year Lung-RADS 1 predictive model are non-nodal specific features.

[0344] For the 5-year Lung-RADS 1 predictive model, the top three features in terms of feature importance are non-nodal specific features. Additionally, the top five features in terms of feature importance for the 5-year Lung-RADS 1 predictive model are non-nodal specific features. Additionally, the top ten features in terms of feature importance for the 5-year Lung-RADS 1 predictive model are non-nodal specific features.

[0345] FIG. 9A shows the performance of the risk prediction models for predicting the likelihood of cancer within 1 year across Lung-RADS 1 patients. The base random forest risk prediction model showed an AUC value of 0.57, whereas the adjusted random forest risk prediction model showed an AUC value of 0.63. FIG. 9B shows the performance of the risk prediction models for predicting the likelihood of cancer within 3 years across Lung-RADS 1 patients. The base random forest risk prediction model showed an AUC value of 0.63, whereas the adjusted random forest risk prediction model showed an AUC value of 0.60. Additionally, Table 16 documents the characteristics of the Lung-RADS 1 patient cohort according to cancer prediction as determined by the 3-year risk prediction models.

[0346] FIG. 9C shows the performance of risk prediction models for predicting the likelihood of cancer within 5 years across Lung-RADS 1 patients. The base random forest risk prediction model showed an AUC value of 0.57, whereas the adjusted random forest risk prediction model showed an AUC value of 0.61. Taken together, the results in FIG. 9A-FIG. 9C show that different risk prediction models can be constructed and deployed to predict the likelihood of cancer in Lung-RADS 1 patient cohorts across different future horizons.

[0347] FIG. 9D shows the 3-year cumulative incidence function across Lung-RADS 1 patients. Here, the risk prediction model predicted cancer for 261 patients and did not predict cancer for the other 4077 patients. Here, the patient cohort includes only Lung-RADS 1 patients, which are the patients at the lowest risk of developing cancer in the near future. As shown in FIG. 9D, the cumulative incidence at the 1-year mark for the 261 predicted cancer patients in the Lung-RADS 1 patient cohort is less than 0.02, while the cumulative incidence at the 1-year mark for the complete patient cohort (e.g., Lung-RADS 1-4B) is approximately 0.2.

[0348] Table 17 shows the enrichment results of the Lung-RADS 1 patient cohort using future cancer predictions from the 1-, 3-, or 5-year risk prediction models. Specifically, Table 17 shows the enrichment results of the risk prediction models compared to the natural incidence of cancer in the original cohort (referred to as the "Null Model" in Table 17). Application of the risk prediction models significantly improves the cumulative incidence (last column of Table 17).

[0349] Specifically, for the 3-year model, the natural incidence rate ("Null model") has a cumulative incidence of 2.31 (e.g., 2.31% of patients in the cohort will be diagnosed with cancer within 3 years). Application of a 3-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence of 5.36 (e.g., the 3-year risk prediction model predicts that 5.36% of patients in the cohort will be diagnosed with cancer within 3 years). Thus, application of the 3-year risk prediction model achieves a 2.3-fold increase in cumulative incidence.

[0350] For the 5-year model, the natural incidence rate ("Null Model") has a cumulative incidence of 3.67 (e.g., 3.67% of patients in the cohort will be diagnosed with cancer within 5 years). Application of a 5-year risk prediction model ("Random Forest") allows for patient enrichment that achieves a cumulative incidence of 6.80 (e.g., the 5-year risk prediction model predicts that 6.80% of patients in the cohort will be diagnosed with cancer within 5 years). Thus, application of the 5-year risk prediction model achieves a 1.9-fold increase in cumulative incidence.

[0351] Taken together, Table 17 shows that various risk prediction models can be implemented to enrich patients in the Lung-RADS 1 cohort, thereby reducing the number of patients that need to be enrolled in clinical trials.

[0352] Example 7: Using radiomics features to predict future risk of cancer for Lung-RADS 1-4B subjects Using the method described in Example 1, two separate models were constructed using Lung-RADS 1-4B subjects. Specifically, the first model is a 1-year, Lung-RADS 1-4B risk prediction model, and the second model is a 3-year, Lung-RADS 1-4B risk prediction model. For each risk prediction model, a training set (e.g., 50% of the cohort) was used to train the risk prediction model, and a test set (e.g., the other 50% of the cohort) was used to test the risk prediction model. The performance of the 1-year Lung-RADS 1-4B and 3-year Lung-RADS 1-4B risk prediction models is shown in Table 18.

[0353] The top 10 important features (e.g., nodal-specific features including radiomics features and non-nodal-specific features) for each of the 1-year and 3-year risk prediction models are shown in Table 19. Notably, for the 1-year Lung-RADS 1-4B prediction model, two of the top three features in terms of feature importance are nodal-specific features. Additionally, three of the top five features in terms of feature importance for the 1-year Lung-RADS 1-4B prediction model are nodal-specific features.

[0354] For the 3-year Lung-RADS 1-4B prediction model, the top three features in terms of feature importance are nodule-specific features.Furthermore, for the 3-year Lung-RADS 1-4B prediction model, the top five features in terms of feature importance are nodule-specific features.

[0355] FIG. 10A shows the performance of a risk prediction model incorporating radiomics features to predict the likelihood of cancer within 1 year across Lung-RADS 1-4B patients. The XGboost risk prediction model showed an AUC value of 0.889. FIG. 10B shows the performance of a risk prediction model incorporating radiomics features to predict the likelihood of cancer within 3 years across Lung-RADS 1-4B patients. The XGboost risk prediction model showed an AUC value of 0.792.

[0356] Example 8: Using radiomics features to predict future risk of cancer for Lung-RADS 1-4A subjects Using the method described in Example 1, two separate models were constructed using Lung-RADS 1-4A subjects. Specifically, the first model is a 1-year, Lung-RADS 1-4A risk prediction model, and the second model is a 3-year, Lung-RADS 1-4A risk prediction model. For each risk prediction model, a training set (e.g., 50% of the cohort) was used to train the risk prediction model, and a test set (e.g., the other 50% of the cohort) was used to test the risk prediction model. The performance of the 1-year Lung-RADS 1-4A and 3-year Lung-RADS 1-4A risk prediction models is shown in Table 18.

[0357] The top 10 important features (e.g., nodal-specific features and non-nodal-specific features, including radiomics features) for each of the 1-year and 3-year risk prediction models are shown in Table 20. In particular, for the 1-year Lung-RADS 1-4A prediction model, the top three features in terms of feature importance are non-nodal-specific features. Additionally, three of the top five features in terms of feature importance for the 1-year Lung-RADS 1-4A prediction model are non-nodal-specific features.

[0358] For the 3-year Lung-RADS 1-4A prediction model, the top three features in terms of feature importance are nodule-specific features.Furthermore, four of the top five features in terms of feature importance for the 3-year Lung-RADS 1-4A prediction model are nodule-specific features.

[0359] FIG. 11A shows the performance of a risk prediction model incorporating radiomics features to predict the likelihood of cancer within 1 year across Lung-RADS 1-4A patients. The XGboost risk prediction model showed an AUC value of 0.779. FIG. 11B shows the performance of a risk prediction model incorporating radiomics features to predict the likelihood of cancer within 3 years across Lung-RADS 1-4A patients. The XGboost risk prediction model showed an AUC value of 0.696.

[0360] Example 9: Using radiomics features to predict future risk of cancer for Lung-RADS 1-3 subjects Using the method described in Example 1, two separate models were constructed using Lung-RADS 1-3 subjects. Specifically, the first model is a 1-year, Lung-RADS 1-3 risk prediction model, and the second model is a 3-year, Lung-RADS 1-3 risk prediction model. For each risk prediction model, a training set (e.g., 50% of the cohort) was used to train the risk prediction model, and a test set (e.g., the other 50% of the cohort) was used to test the risk prediction model. The performance of the 1-year Lung-RADS 1-3 and 3-year Lung-RADS 1-3 risk prediction models is shown in Table 18.

[0361] The top 10 important features (e.g., nodal-specific features and non-nodal-specific features, including radiomics features) for each of the 1-year and 3-year risk prediction models are shown in Table 21. In particular, for the 1-year Lung-RADS 1-3 prediction model, the top three features in terms of feature importance are non-nodal-specific features. Additionally, three of the top five features in terms of feature importance for the 1-year Lung-RADS 1-3 prediction model are non-nodal-specific features.

[0362] For the 3-year Lung-RADS 1-3 predictive model, two of the top three features in terms of feature importance are non-nodal specific features.Furthermore, three of the top five features in terms of feature importance for the 3-year Lung-RADS 1-3 predictive model are non-nodal specific features.

[0363] FIG. 12A shows the performance of a risk prediction model incorporating radiomics features to predict the likelihood of cancer within 1 year across Lung-RADS 1-3 patients. The XGboost risk prediction model showed an AUC value of 0.793. FIG. 12B shows the performance of a risk prediction model incorporating radiomics features to predict the likelihood of cancer within 3 years across Lung-RADS 1-3 patients. The XGboost risk prediction model showed an AUC value of 0.676.

[0364] Example 10: Using radiomics features to predict future risk of cancer for Lung-RADS 2-4B subjects Using the method described in Example 1, two separate models were constructed using Lung-RADS 2-4B subjects. Specifically, the first model is a 1-year, Lung-RADS 2-4B risk prediction model, and the second model is a 3-year, Lung-RADS 2-4B risk prediction model. For each risk prediction model, a training set (e.g., 50% of the cohort) was used to train the risk prediction model, and a test set (e.g., the other 50% of the cohort) was used to test the risk prediction model. The performance of the 1-year Lung-RADS 2-4B and 3-year Lung-RADS 2-4B risk prediction models is shown in Table 18.

[0365] The top 10 important features (e.g., nodal-specific features and non-nodal-specific features, including radiomics features) for each of the 1-year and 3-year risk prediction models are shown in Table 22. In particular, for the 1-year Lung-RADS 2-4B prediction model, the top three features in terms of feature importance are nodal-specific features. Additionally, the top five features in terms of feature importance for the 1-year Lung-RADS 2-4B prediction model are nodal-specific features.

[0366] For the 3-year Lung-RADS 2-4B risk prediction model, the top three features in terms of feature importance are nodule-specific features.Furthermore, for the 3-year Lung-RADS 2-4B risk prediction model, the top five features in terms of feature importance are nodule-specific features.

[0367] FIG. 13A shows the performance of a risk prediction model incorporating radiomics features to predict the likelihood of cancer within 1 year across Lung-RADS 2-4B patients. The XGboost risk prediction model showed an AUC value of 0.928. FIG. 13B shows the performance of a risk prediction model incorporating radiomics features to predict the likelihood of cancer within 3 years across Lung-RADS 2-4B patients. The XGboost risk prediction model showed an AUC value of 0.809.

[0368] Example 11: Using radiomics features to predict future risk of cancer in Lung-RADS 4A-4B subjects Using the method described in Example 1, two separate models were constructed using Lung-RADS 4A-4B subjects. Specifically, the first model is a 1-year, Lung-RADS 4A-4B risk prediction model, and the second model is a 3-year, Lung-RADS 4A-4B risk prediction model. For each risk prediction model, a training set (e.g., 50% of the cohort) was used to train the risk prediction model, and a test set (e.g., the other 50% of the cohort) was used to test the risk prediction model. The performance of the 1-year Lung-RADS 4A-4B and 3-year Lung-RADS 4A-4B risk prediction models is shown in Table 18.

[0369] The top 10 important features (e.g., nodal-specific features including radiomics features and non-nodal-specific features) for each of the 1-year and 3-year risk prediction models are shown in Table 23. In particular, for the 1-year Lung-RADS 4A-4B prediction model, the top three features in terms of feature importance are nodal-specific features. Additionally, the top five features in terms of feature importance for the 1-year Lung-RADS 4A-4B prediction model are nodal-specific features.

[0370] For the 3-year Lung-RADS 4A-4B risk prediction model, the top three features in terms of feature importance are nodule-specific features.Furthermore, for the 3-year Lung-RADS 4A-4B risk prediction model, the top five features in terms of feature importance are nodule-specific features.

[0371] FIG. 14A shows the performance of a risk prediction model incorporating radiomics features to predict the likelihood of cancer within 1 year across Lung-RADS 4A-4B patients. The XGboost risk prediction model showed an AUC value of 0.91. FIG. 14B shows the performance of a risk prediction model incorporating radiomics features to predict the likelihood of cancer within 3 years across Lung-RADS 4A-4B patients. The XGboost risk prediction model showed an AUC value of 0.798.

[0372] Example 12: Additional Examples of Predicting Future Risk of Cancer in Lung-RADS 1-4B Subjects Additional models are constructed using Lung-RADS 1-4B subjects using the methods described in Example 1. For example, such models can incorporate the non-nodule features described in Example 1, as well as the nodule-specific features (such as radiomics features) described in Example 1.

[0373] The additional models are either random forest models or gradient boosting models. Specifically, the first model is a 6-month Lung-RADS 1-4B risk prediction model, the second model is a 1.5-year Lung-RADS 1-4B risk prediction model, the third model is a 2.5-year Lung-RADS 1-4B risk prediction model, the fourth model is a 4-year Lung-RADS 1-4B risk prediction model, the fifth model is a 4.5-year Lung-RADS 1-4B risk prediction model, the sixth model is a 5.5-year Lung-RADS 1-4B risk prediction model, the seventh model is a 6-year Lung-RADS 1-4B risk prediction model, the eighth model is a 7-year Lung-RADS 1-4B risk prediction model, the ninth model is an 8-year Lung-RADS 1-4B risk prediction model, and the tenth model is a 9-year Lung-RADS 1-4B risk prediction model. the 11th model is a 10-year Lung-RADS 1-4B risk prediction model, the 12th model is an 11-year Lung-RADS 1-4B risk prediction model, the 13th model is an 11-year Lung-RADS 1-4B risk prediction model, the 14th model is a 12-year Lung-RADS 1-4B risk prediction model, the 15th model is a 13-year Lung-RADS 1-4B risk prediction model, the 16th model is a 14-year Lung-RADS 1-4B risk prediction model, the 17th model is a 15-year Lung-RADS 1-4B risk prediction model, the 18th model is a 16-year Lung-RADS 1-4B risk prediction model, and the 19th model is a 17-year Lung-RADS The 20th model is a 18-year Lung-RADS 1-4B risk prediction model, the 21st model is a 19-year Lung-RADS 1-4B risk prediction model, and the 22nd model is a 20-year Lung-RADS 1-4B risk prediction model.

[0374] The additional models exhibit an AUC value of at least 0.55. At least one of the additional models exhibits an AUC value of at least 0.60. At least one of the additional models exhibits an AUC value of at least 0.65. At least one of the additional models exhibits an AUC value of at least 0.70. At least one of the additional models exhibits an AUC value of at least 0.75. At least one of the additional models exhibits an AUC value of at least 0.80.

[0375] Example 13: Additional Examples of Predicting Future Risk of Cancer for Lung-RADS 1-4A Subjects Additional models are constructed using Lung-RADS 1-4A subjects using the methods described in Example 1. For example, such models can incorporate the non-nodule features described in Example 1, as well as the nodule-specific features (such as radiomics features) described in Example 1.

[0376] The additional models are either random forest models or gradient boosting models. Specifically, the first model is a 6-month Lung-RADS 1-4A risk prediction model, the second model is a 1.5-year Lung-RADS 1-4A risk prediction model, the third model is a 2.5-year Lung-RADS 1-4A risk prediction model, the fourth model is a 4-year Lung-RADS 1-4A risk prediction model, the fifth model is a 4.5-year Lung-RADS 1-4A risk prediction model, the sixth model is a 5.5-year Lung-RADS 1-4A risk prediction model, the seventh model is a 6-year Lung-RADS 1-4A risk prediction model, the eighth model is a 7-year Lung-RADS 1-4A risk prediction model, the ninth model is an 8-year Lung-RADS 1-4A risk prediction model, and the tenth model is a 9-year Lung-RADS 1-4A risk prediction model. the 11th model is a 10-year Lung-RADS 1-4A risk prediction model, the 12th model is an 11-year Lung-RADS 1-4A risk prediction model, the 13th model is an 11-year Lung-RADS 1-4A risk prediction model, the 14th model is a 12-year Lung-RADS 1-4A risk prediction model, the 15th model is a 13-year Lung-RADS 1-4A risk prediction model, the 16th model is a 14-year Lung-RADS 1-4A risk prediction model, the 17th model is a 15-year Lung-RADS 1-4A risk prediction model, the 18th model is a 16-year Lung-RADS 1-4A risk prediction model, and the 19th model is a 17-year Lung-RADS The first model is a 18-year Lung-RADS 1-4A risk prediction model, the second model is a 19-year Lung-RADS 1-4A risk prediction model, and the third model is a 20-year Lung-RADS 1-4A risk prediction model.

[0377] The additional models exhibit an AUC value of at least 0.55. At least one of the additional models exhibits an AUC value of at least 0.60. At least one of the additional models exhibits an AUC value of at least 0.65. At least one of the additional models exhibits an AUC value of at least 0.70. At least one of the additional models exhibits an AUC value of at least 0.75. At least one of the additional models exhibits an AUC value of at least 0.80.

[0378] Example 14: Additional Examples of Predicting Future Risk of Cancer in Lung-RADS 1-3 Subjects Additional models are constructed using Lung-RADS 1-3 subjects using the methods described in Example 1. For example, such models can incorporate the non-nodule features described in Example 1, as well as the nodule-specific features (such as radiomics features) described in Example 1.

[0379] The additional models are either random forest models or gradient boosting models. Specifically, the first model is a 6-month Lung-RADS 1-3 risk prediction model, the second model is a 1.5-year Lung-RADS 1-3 risk prediction model, the third model is a 2.5-year Lung-RADS 1-3 risk prediction model, the fourth model is a 4-year Lung-RADS 1-3 risk prediction model, the fifth model is a 4.5-year Lung-RADS 1-3 risk prediction model, the sixth model is a 5.5-year Lung-RADS 1-3 risk prediction model, the seventh model is a 6-year Lung-RADS 1-3 risk prediction model, the eighth model is a 7-year Lung-RADS 1-3 risk prediction model, the ninth model is an 8-year Lung-RADS 1-3 risk prediction model, and the tenth model is a 9-year Lung-RADS 1-3 risk prediction model. The 11th model is a 10-year Lung-RADS 1-3 risk prediction model, the 12th model is a 11-year Lung-RADS 1-3 risk prediction model, the 13th model is a 11-year Lung-RADS 1-3 risk prediction model, the 14th model is a 12-year Lung-RADS 1-3 risk prediction model, the 15th model is a 13-year Lung-RADS 1-3 risk prediction model, the 16th model is a 14-year Lung-RADS 1-3 risk prediction model, the 17th model is a 15-year Lung-RADS 1-3 risk prediction model, the 18th model is a 16-year Lung-RADS 1-3 risk prediction model, the 19th model is a 17-year Lung-RADS 1-3 risk prediction model, and the 20th model is an 18-year Lung-RADS The 21st model is a 19-year Lung-RADS 1-3 risk prediction model, and the 22nd model is a 20-year Lung-RADS 1-3 risk prediction model.

[0380] The additional models exhibit an AUC value of at least 0.55. At least one of the additional models exhibits an AUC value of at least 0.60. At least one of the additional models exhibits an AUC value of at least 0.65. At least one of the additional models exhibits an AUC value of at least 0.70. At least one of the additional models exhibits an AUC value of at least 0.75. At least one of the additional models exhibits an AUC value of at least 0.80.

[0381] Example 15: Additional Examples of Predicting Future Risk of Cancer in Lung-RADS 1-2 Subjects Additional models are constructed using Lung-RADS 1-2 subjects using the methods described in Example 1. For example, such models can incorporate the non-nodule features described in Example 1, as well as the nodule-specific features (such as radiomics features) described in Example 1.

[0382] The additional models are either random forest models or gradient boosting models. Specifically, the first model is a 6-month Lung-RADS 1-2 risk prediction model, the second model is a 1.5-year Lung-RADS 1-2 risk prediction model, the third model is a 2.5-year Lung-RADS 1-2 risk prediction model, the fourth model is a 4-year Lung-RADS 1-2 risk prediction model, the fifth model is a 4.5-year Lung-RADS 1-2 risk prediction model, the sixth model is a 5.5-year Lung-RADS 1-2 risk prediction model, the seventh model is a 6-year Lung-RADS 1-2 risk prediction model, the eighth model is a 7-year Lung-RADS 1-2 risk prediction model, the ninth model is an 8-year Lung-RADS 1-2 risk prediction model, and the tenth model is a 9-year Lung-RADS 1-2 risk prediction model. the 11th model is a 10-year Lung-RADS 1-2 risk prediction model, the 12th model is a 11-year Lung-RADS 1-2 risk prediction model, the 13th model is a 11-year Lung-RADS 1-2 risk prediction model, the 14th model is a 12-year Lung-RADS 1-2 risk prediction model, the 15th model is a 13-year Lung-RADS 1-2 risk prediction model, the 16th model is a 14-year Lung-RADS 1-2 risk prediction model, the 17th model is a 15-year Lung-RADS 1-2 risk prediction model, the 18th model is a 16-year Lung-RADS 1-2 risk prediction model, the 19th model is a 17-year Lung-RADS 1-2 risk prediction model, and the 20th model is an 18-year Lung-RADS the 21st model is a 19-year Lung-RADS 1-2 risk prediction model, and the 22nd model is a 20-year Lung-RADS 1-2 risk prediction model.

[0383] The additional models exhibit an AUC value of at least 0.55. At least one of the additional models exhibits an AUC value of at least 0.60. At least one of the additional models exhibits an AUC value of at least 0.65. At least one of the additional models exhibits an AUC value of at least 0.70. At least one of the additional models exhibits an AUC value of at least 0.75. At least one of the additional models exhibits an AUC value of at least 0.80.

[0384] Example 15: Additional Examples of Predicting Future Risk of Cancer in Lung-RADS 1 Subjects Additional models are constructed using the methods described in Example 1 and using Lung-RADS 1 subjects. For example, such models can incorporate the non-nodule features described in Example 1, and the nodule-specific features (such as radiomics features) described in Example 1.

[0385] The additional models are either random forest models or gradient boosting models. Specifically, the first model is a 6-month Lung-RADS 1 risk prediction model, the second model is a 1.5-year Lung-RADS 1 risk prediction model, the third model is a 2.5-year Lung-RADS 1 risk prediction model, the fourth model is a 4-year Lung-RADS 1 risk prediction model, the fifth model is a 4.5-year Lung-RADS 1 risk prediction model, the sixth model is a 5.5-year Lung-RADS 1 risk prediction model, the seventh model is a 6-year Lung-RADS 1 risk prediction model, the eighth model is a 7-year Lung-RADS 1 risk prediction model, the ninth model is an 8-year Lung-RADS 1 risk prediction model, the tenth model is a 9-year Lung-RADS 1 risk prediction model, and the eleventh model is a 10-year Lung-RADS 1 risk prediction model. 1 risk prediction model, the 12th model is an 11-year Lung-RADS 1 risk prediction model, the 13th model is an 11-year Lung-RADS 1 risk prediction model, the 14th model is a 12-year Lung-RADS 1 risk prediction model, the 15th model is a 13-year Lung-RADS 1 risk prediction model, the 16th model is a 14-year Lung-RADS 1 risk prediction model, the 17th model is a 15-year Lung-RADS 1 risk prediction model, the 18th model is a 16-year Lung-RADS 1 risk prediction model, the 19th model is a 17-year Lung-RADS 1 risk prediction model, the 20th model is an 18-year Lung-RADS 1 risk prediction model, the 21st model is a 19-year Lung-RADS 1 risk prediction model, and the 22nd model is a 20-year Lung-RADS 1 Risk prediction model.

[0386] The additional models exhibit an AUC value of at least 0.55. At least one of the additional models exhibits an AUC value of at least 0.60. At least one of the additional models exhibits an AUC value of at least 0.65. At least one of the additional models exhibits an AUC value of at least 0.70. At least one of the additional models exhibits an AUC value of at least 0.75. At least one of the additional models exhibits an AUC value of at least 0.80.

[0387] Example 16: Example of future risk prediction for two patients Risk prediction models trained using data from the NLST provide a multimodal risk assessment and enrichment approach. For example, consider two patients who developed cancer within 3 years of enrollment in the NLST. Both had Lung-RADS 3 nodules at baseline, both were predicted by the algorithm to develop cancer within 3 years, and both were predicted to develop IA adenocarcinoma within 3 years. In fact, the predicted probability of developing cancer for both patients is within 2% of each other (84.8% for patient 1 and 86.4% for patient 2). Patient 1 was a 65-year-old white female ever smoker with a 76 pack-year smoking history, a BMI of 23.5, relatively small pectoral muscle mass, and significant upper lobe predominant emphysema (31% emphysema based on regional histogram analysis). A 76 pack-year smoking history is equivalent to smoking one pack (20 cigarettes) of cigarettes daily for 76 years or two packs daily for 38 years. In contrast, patient 2 was a 59-year-old white male ever smoker with a 43 pack-year smoking history, a BMI of 29.8, relatively preserved pectoral muscle mass, and minimal emphysema (2.6% emphysema by volume based on local histogram analysis), but significantly more interstitial features (7.1% compared to 4%). In this second case, this last feature and the strong relationship between stromal features and cancer likely led the algorithm to predict cancer development within 3 years. The clinical and radiological differences between these two cases highlight the ability of risk prediction models to identify patients with unique and distinct phenotypes likely to develop cancer.

[0388] All publications, patents, patent applications, and other documents cited in this application are incorporated by reference in their entirety for all purposes to the same extent as if each individual publication, patent, patent application, and other document was individually indicated to be incorporated by reference for all purposes.

[0389] While various specific embodiments have been illustrated and described, the above specification is not limiting. It will be recognized that various modifications can be made without departing from the spirit and scope of the disclosure. Many variations will become apparent to those skilled in the art upon review of this specification.

[0390] table

[0391] [Table 1]

[0392] [Table 2]

[0393] [Table 3]

[0394] [Table 4-1]

[0395] [Table 4-2]

[0396] [Table 5]

[0397] [Table 6]

[0398]

Table 7-1

[0399]

Table 7-2

[0400]

Table 8

[0401]

Table 9

[0402]

Table 10-1

[0403]

Table 10-2

[0404]

Table 11

[0405]

Table 12

[0406]

Table 13-1

[0407]

Table 13-2

[0408]

Table 14

[0409]

Table 15

[0410]

Table 16-1

[0411]

Table 16-2

[0412]

Table 17

[0413]

Table 18

[0414]

Table 19

[0415]

Table 20

[0416]

Table 21

[0417]

Table 22

[0418] [Table 23]

[0419] [Embodiment] (1) A method for predicting one or more future risks of lung cancer in a subject, comprising: obtaining one...

Claims

1. 1. A method for predicting one or more future risks of lung cancer in a subject, comprising: obtaining one or more images captured of the object at a single time; extracting features from the one or more acquired images, the extracted features including at least non-nodule specific features, the non-nodule specific features including both lung parenchyma features and body composition features, the body composition features including pectoral muscle or subcutaneous fat features of the subject; and predicting one or more future risks of lung cancer for the subject by applying one or more trained risk prediction models to analyze the features extracted from the one or more acquired images.

2. The method of claim 1 , wherein the lung parenchyma features include one or more of densitometry measurements of the lung parenchyma or local histogram measurements of the lung parenchyma.

3. The method of claim 1, wherein the characteristics of the pectoral muscle or subcutaneous fat include pectoral muscle cross-sectional area, pectoral muscle lean cross-sectional area, or subcutaneous fat cross-sectional area.

4. 2. The method of claim 1, wherein the extracted features further comprise nodule-specific features selected from one or more of nodule-specific attenuation, nodule margin description, nodule size, nodule shape, nodule texture, nodule diameter, Lung-RADS score, or radiomics features.

5. 2. The method of claim 1, wherein predicting the one or more future risks of lung cancer in the subject comprises applying an M risk prediction model to predict whether the subject is likely to develop lung cancer within M.

6. 2. The method of claim 1, wherein one of the one or more risk prediction models is trained to predict future risk of lung cancer using training images captured from a training cohort that includes training individuals classified into any of Lung-RADS 1-4B, training individuals classified into any of Lung-RADS 1-4A, training individuals classified into any of Lung-RADS 1-3, training individuals classified into any of Lung-RADS 1-2, or training individuals classified into any of Lung-RADS 1.

7. The method of claim 1 , wherein the one or more images include a chest CT image or a chest X-ray image.

8. prior to predicting one or more future risks of lung cancer for said subject, obtaining nodule-specific features corresponding to said subject; 10. The method of claim 1, further comprising determining that the subject is a candidate for future risk prediction based on the nodule-specific features.

9. 10. The method of claim 1, further comprising selecting a clinical response for the subject based on the predicted future risk of lung cancer.

10. A non-transitory computer-readable medium for predicting one or more future risks of lung cancer in a subject, the non-transitory computer-readable medium comprising instructions that, when executed by a processor, cause the processor to: obtaining one or more images captured of the object at a single time; extracting features from the one or more acquired images, the extracted features including at least non-nodule specific features, the non-nodule specific features including both lung parenchyma features and body composition features, the body composition features including pectoral muscle or subcutaneous fat features of the subject; and applying one or more trained risk prediction models to predict one or more future risks of lung cancer for the subject by analyzing the features extracted from the one or more acquired images.

Images