Breast cancer 70 gene detection risk assessment method based on image and pathology

Abstract

The invention relates to the field of tumor prognosis evaluation, and discloses an image and pathology-based breast cancer 70 gene detection risk evaluation method, which comprises the following steps of: selecting an HR < + > / HER2-early breast cancer patient meeting a condition as a research object, and acquiring an ultrasonic image and clinical pathology information of the HR < + > / HER2-early breast cancer patient; the method comprises the following steps: carrying out ROI (Region of Interest) sketching on a tumor region through ImageJ software, and extracting image omics characteristics by utilizing a Pyradiomics tool; after z-score standardization is carried out on the features, LASSO regression is adopted to carry out feature screening, and a radiomics score and a tumor size are calculated; combining clinical parameters such as age, lymph node metastasis number, ER, PR, Her-2, Ki67, histological grading, vessel invasion and the like to construct a column graph model; a recurrence risk probability value is further calculated by using a linear predictor, and the patients are classified into high-risk or low-risk categories. By integrating multi-dimensional structured data, a recurrence risk prediction model with excellent performance and high clinical applicability is established.

Description

Technical Field

[0001] The present invention relates to the technical field of tumor prognosis assessment, and in particular to a method for risk assessment of breast cancer 70 genes based on imaging and pathology. Background Art

[0002] Breast cancer is one of the most common malignant tumors among women worldwide. In the selection of treatment strategies, how to accurately identify the risk of recurrence and optimize adjuvant chemotherapy decisions are key links in improving prognosis and reducing the burden of treatment. According to the NCCN (National Comprehensive Cancer Network) guidelines, the 70-gene expression profile test (MammaPrint, abbreviated as 70-GS) can be used to stratify the risk of distant recurrence in patients with HR-positive / HER2-negative early breast cancer, especially for patients with negative lymph nodes or 1-3 positive lymph nodes. Through the stratification mechanism established by 70-GS, doctors can more scientifically judge whether patients benefit from chemotherapy and the extent of benefit, thereby achieving more personalized treatment plans.

[0003] In actual clinical applications, 70-GS can not only be used to quantitatively predict the 5-year and 10-year recurrence risk, but also serve as a decision-making support tool in the adjuvant treatment pathway. For example, the results of the MINDACT study showed that among patients with high clinical risk but low genetic risk, a considerable proportion can avoid chemotherapy, effectively reducing the toxic side effects and quality of life damage caused by overtreatment. At the same time, the risk classification of 70-GS was also found to be correlated with the pathological complete response rate (pCR) after neoadjuvant therapy, and has the dual value of predicting treatment sensitivity and long-term prognosis.

[0004] While the accuracy and guiding significance of 70-GS have been extensively validated by evidence-based studies, its widespread adoption in clinical practice remains challenging. The most prominent limitation is the high cost of testing. The current international price of 70-GS remains prohibitively expensive for families in many developing countries. In China, the cost of a single test can reach 50%-75% of the average annual income, making this high-throughput genetic testing inaccessible to most patients and hindering the equitable implementation of gene-guided treatment strategies.

[0005] At the same time, while traditional clinical parameters and imaging features can be used to assess tumor malignancy and recurrence propensity to a certain extent, they are highly subjective, have poor quantitative capabilities, and are unable to deeply reflect the biological heterogeneity within the tumor, limiting their potential for accurate stratification of recurrence risk. Therefore, there is an urgent need to develop a risk prediction method based on non-invasive data, which is low-cost and reproducible, to replace or supplement existing genetic testing technologies and thus expand the application boundaries of precision medicine strategies.

[0006] The present invention is proposed precisely to address the above-mentioned issues, aiming to construct a recurrence risk prediction model that integrates ultrasound imaging genomics features with clinical pathological information, replacing expensive genetic testing methods and providing breast cancer patients with an economically accessible and high-performance risk assessment tool, thereby enhancing the scientific nature of auxiliary treatment decisions, reducing unnecessary treatments, and ultimately improving patient prognosis and quality of life. Summary of the Invention

[0007] The purpose of this invention is to provide a 70-gene detection risk assessment method for breast cancer based on imaging and pathology, so as to solve the technical problem of relying on high-cost gene detection methods in existing breast cancer recurrence risk assessment.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: The first aspect of the present invention provides a method for risk assessment of breast cancer 70 genes based on imaging and pathology, comprising the following steps: 1) Obtain clinicopathological parameters of patients with HR+ / HER2- early breast cancer; 2) Obtain the patient's breast ultrasound image, delineate the lesion area, and extract radiomics features; 3) LASSO regression was used to screen target feature variables from the extracted radiomics features and construct radiomics scores; 4) Combining clinical pathological parameters with radiomics scores, a nomogram prediction model was constructed, and each parameter was mapped into a quantitative score to obtain a total score; 5) Based on the total score, the risk probability value is calculated by a linear predictor to classify the patients into high-risk or low-risk categories.

[0009] Preferably, the clinical pathological parameters include: Patient age, number of lymph node metastases, estrogen receptor status, progesterone receptor status, human epidermal growth factor receptor-2 status, Ki67 index, histological grade, and vascular invasion; The clinical pathological parameters were collected before the patients underwent breast ultrasound examination, and the patients did not receive neoadjuvant chemotherapy, radiotherapy, puncture biopsy or local breast surgery.

[0010] Preferably, the breast ultrasound image is acquired by a medical ultrasound device with high-resolution imaging capability, and the lesion image is imaged by a professional imaging physician for ROI delineation; The consistency of the ROI delineation is evaluated by the intra-group correlation coefficient. Only when the intra-group correlation coefficient is higher than a preset consistency threshold, the image data enters the subsequent feature extraction step.

[0011] Preferably, the radiomics features are extracted by an image analysis tool, and the radiomics features include: Shape characteristics; Grayscale histogram features; Gray-level co-occurrence matrix features; Grayscale run length matrix features; Grayscale size area matrix features; Neighborhood grayscale difference matrix features; After the radiomics features were normalized by Z-score, LASSO regression was used for variable screening, and collinear variables were eliminated by variance inflation factor calculation. Only features that met the preset collinearity threshold were retained for model construction.

[0012] Preferably, the radiomics score is obtained by a radiomics scoring model, and the radiomics scoring model is constructed based on multiple significant feature variables screened by LASSO regression, and the variables are respectively from one or more of the following categories: shape features; Grayscale histogram features; Texture features, including GLCM, GLRLM, GLSZM and NGTDM features; The score is obtained by multiplying the value of each variable by its corresponding regression coefficient and performing weighted summation to form a quantitative Radiomics score, which is used to construct a nomogram model.

[0013] Preferably, the correlation between the target feature variables screened out from the imaging omics features and the patient's 70-gene expression profile risk classification results is verified by Spearman rank correlation coefficient analysis, and has a statistically significant positive or negative correlation.

[0014] Preferably, the nomogram prediction model is constructed using a logistic regression method, wherein the Radiomics score is used as one of the candidate modeling variables, and specifically comprises the following steps: 1) Input candidate predictor variables into the logistic regression model and perform multivariate stepwise regression analysis; 2) converting modeling variables into a quantitative scoring system and accumulating the patient's total score based on it; 3) Based on the logistic regression linear prediction function, the total score is converted into a predicted probability value, and a nomogram is drawn to achieve individualized risk prediction.

[0015] Preferably, the nomogram model is constructed in the following three versions for comparative verification: Radiomics models constructed based only on radiomics scores and tumor size; Clinical models constructed based solely on clinicopathological parameters; A joint imaging-clinical model constructed by integrating imaging scores and clinical parameters; The imaging omics model, clinical model and imaging-clinical joint model are all constructed based on the same sample data set.

[0016] Preferably, the risk probability value output by the model is a continuous value between 0 and 1; Based on empirical judgment criteria, when the probability value is higher than the median value, the patient is judged to be in the high-risk category of the 70-gene breast cancer test and is suitable for subsequent intensive treatment; When the probability value is lower than the median value, it is judged as a low-risk category, and unnecessary chemotherapy intervention can be avoided.

[0017] A second aspect of the present invention provides a breast cancer risk assessment system for implementing the method, comprising: An image acquisition module, used for acquiring breast ultrasound images; Pathology information acquisition module, used to obtain the patient's clinical pathology parameters; Feature extraction module, which extracts radiomic features based on the ROI region of the breast image; The risk modeling module uses radiomics scores and clinical parameters to construct a nomogram model; The risk output module outputs high / low risk assessment results based on the total score.

[0018] In summary, the present invention includes at least one of the following beneficial technical effects: 1. This paper combines radiomics scores with tumor size and uses the LASSO regression algorithm to extract and screen multiple stable and predictive radiomics features from ultrasound images. This method achieves deep quantitative expression of breast cancer tumor morphology and texture information, thereby effectively modeling the biological behavior of tumors without relying on invasive testing, providing a structured quantitative basis for predicting the risk of distant recurrence.

[0019] 2. The present invention introduces multivariate logistic regression analysis, integrates clinical pathological parameters (including age, number of lymph node metastases, ER, PR, Her-2 status, Ki67, histological grade and vascular invasion) with imaging genomics scoring variables, and constructs an imaging-clinical joint nomogram model. This model has the ability to express the characteristics of both morphological and clinical biomarker dimensions, can more comprehensively reflect the recurrence risk characteristics of individual patients, and achieve complementarity in risk assessment dimensions.

[0020] 3. This invention uses a nomogram modeling approach to map high-dimensional imaging genomics and clinical parameters into a visual scoring system, and outputs the recurrence risk probability based on a linear predictor, thereby achieving an intuitive display and quantitative expression of the model results. This helps doctors quickly interpret the model output in clinical practice, thereby improving the model's interpretability and practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the method flow of the present invention; Figure 2 A schematic diagram of a nomogram model constructed according to the present invention; Figure 3 A radiomics model constructed for an embodiment of the present invention; Figure 4 A clinical model constructed for an embodiment of the present invention; Figure 5 The imaging-clinical joint model constructed for the embodiment of the present invention; Figure 6 ROC curves of the three models constructed in the embodiments of the present invention; (a) (d) are ROC curves of the combined imaging-clinical model in training set a and validation set d; (b) (e) are ROC curves of the clinical model in training set b and validation set e; (c) (f) are ROC curves of the imaging omics model in training set c and validation set f; Figure 7 AUC comparison chart of the three models constructed in the embodiment of the present invention on the training set and the validation set; Figure 8 Precision-recall (PR) curves of the three models constructed according to the present invention; (a) and (d) are the PR curves of the combined imaging-clinical model in training set a and validation set d; (b) and (e) are the PR curves of the clinical model in training set b and validation set e; (c) and (f) are the PR curves of the imaging omics model in training set c and validation set f. Figure 9The figures show the calibration curves and model cost-effectiveness analysis of the three models constructed in accordance with the present invention. The model calibration curves (af) are generated using the bootstrap method with 500 repetitions, and the apparent, bias-corrected, and ideal lines are estimated using a resampling technique based on the sample size n (training or validation set size). (a) and (d) are the calibration curves of the combined imaging-clinical model in training set a and validation set d. (b) and (e) are the calibration curves of the clinical model in training set b and validation set e. (c) and (f) are the calibration curves of the imaging omics model in training set c and validation set f. In the model cost-effectiveness analysis (gh), (g) and (h) are the standardized net benefit curves of the three models in training set g and validation set h, as well as their performance at different high-risk thresholds and cost-effectiveness ratios. The vertical axis represents the standardized net benefit, and the horizontal axis represents the high-risk threshold (0-1), indicating the cutoff probability at which the model classifies a patient as high-risk. The cost-effectiveness ratio is plotted below the horizontal axis, reflecting the trade-off between the cost of false positives and the benefit of true positives.

[0022] Figure 10 Schematic diagram of the system structure of the present invention.

[0023] Among them, 10, image acquisition module; 20, pathology information acquisition module; 30, feature extraction module; 40, risk modeling module; 50, risk output module. DETAILED DESCRIPTION

[0024] The following is combined with Figure 1 -Attached Figure 10 , the present invention is described in further detail.

[0025] The present invention provides a 70-gene breast cancer risk assessment method based on imaging and pathology. This method constructs a nomogram prediction model by combining multimodal data to replace or supplement high-cost genetic testing and achieve accurate stratification of patient recurrence risk.

[0026] like Figure 1 As shown, the imaging and pathology-based breast cancer 70-gene detection risk assessment method may include the following steps: S1. Obtain clinicopathological parameters of patients with HR+ / HER2- early breast cancer; S2. Obtain the patient's breast ultrasound image, delineate the lesion area, and extract radiomics features; S3. Use LASSO regression to select target feature variables from the extracted radiomics features, determine tumor size based on the radiomics features, and construct a radiomics score; S4. Combining clinical pathological parameters with radiomics scores, a nomogram prediction model was constructed, and each parameter was mapped into a quantitative score to obtain a total score; S5. Calculate the risk probability value based on the total score using a linear predictor to classify the patient into a high-risk or low-risk category.

[0027] The following is a detailed description of each step in the method of the present invention, which comprehensively explains the specific implementation principles, technical details and processes of each step.

[0028] In step S1, obtaining the clinicopathological parameters of patients with HR+ / HER2- early breast cancer provides key clinical background information for subsequent risk assessment and ensures that the collected data can accurately reflect the individual characteristics of patients during the model construction process.

[0029] In the method of the present invention, the collection of clinical pathological parameters is crucial, as they can provide basic data support for predicting the recurrence risk of breast cancer patients. To ensure the accuracy and reliability of the data, all clinical pathological parameters must be collected before the patient undergoes breast ultrasound examination. Specifically, clinical pathological parameters include but are not limited to: Patient age: Age is an important factor in the prognosis of breast cancer patients. Obtaining patient age information can provide important biological background data for risk assessment models.

[0030] Number of lymph node metastases: Lymph node metastasis is one of the key indicators in the clinical evaluation of breast cancer. The number of lymph node metastases is directly related to the staging and treatment decision of breast cancer. Therefore, in this embodiment, the number of lymph node metastases of the patient needs to be recorded in detail.

[0031] ER, PR, and Her-2 status: These hormone receptors and human epidermal growth factor receptor 2 (HER2) status are important for breast cancer classification, treatment selection, and prognosis assessment. ER and PR status can help determine whether a patient is suitable for hormone therapy, while Her-2 status is related to whether targeted therapy can be used.

[0032] Ki67 index: The Ki67 index reflects the proliferation activity of tumor cells and is often used to assess tumor invasiveness. The Ki67 index is an important indicator for breast cancer prognosis assessment.

[0033] Histological grade: Histological grade is an evaluation of the morphological characteristics of tumor tissues through pathological examination, which is used to measure the degree of malignancy of the tumor. Higher histological grade is generally associated with a worse prognosis.

[0034] Vascular invasion: Vascular invasion is a key pathway for breast cancer metastasis. Detection of vascular invasion reflects the aggressiveness and potential for metastasis of a tumor, and therefore plays a crucial role in clinical pathological parameters.

[0035] All of these clinical and pathological parameters are collected by medical staff before the patient undergoes a breast ultrasound examination to ensure that the data provided is representative and accurately reflects the patient's clinical condition. The acquisition of these parameters must strictly follow standardized operating procedures to ensure consistency and reliability of the data.

[0036] Furthermore, when collecting clinical pathology parameters, patients must not have received neoadjuvant chemotherapy, radiotherapy, biopsy, or local breast surgery. This requirement is intended to eliminate the interference of external treatment factors on the data results and ensure that the collected parameter data reflects the patient's true biological characteristics, thereby improving the accuracy of subsequent risk assessment models.

[0037] In this example, the collection of clinical pathological parameters not only provides important reference data for the 70-gene test for breast cancer but also lays the foundation for the integration of radiomics features with clinical information. By combining these clinical pathological parameters with subsequently extracted radiomics features, the method of the present invention can construct a more comprehensive and accurate risk assessment model, providing more personalized treatment plans for breast cancer patients.

[0038] In step S2, a breast ultrasound image of the patient to be evaluated is first acquired, and a region of interest (ROI) is delineated for the lesion area. The purpose of this step is to extract effective radiomics features from the image data, which will provide key information for the subsequent risk assessment model.

[0039] In this embodiment, breast ultrasound images were acquired using high-resolution medical ultrasound equipment. Preferred ultrasound equipment includes, but is not limited to, the Samsung WS80A, Philips EPIQ 5&7, Philips IU22, and GE Logic 9. These devices offer high spatial resolution and image quality, clearly displaying breast tissue details, particularly the size, shape, and margin characteristics of tumors. The breast ultrasound images acquired by these devices provide precise imaging data for subsequent analysis.

[0040] Subsequently, the imaging specialist will accurately outline the lesion area in the image. ROI outline refers to manually marking the outline of the tumor based on the tumor area in the image. The accuracy of ROI outline is crucial to the subsequent image feature extraction, so it needs to be performed by specially trained imaging experts. This outline process requires precise depiction of the boundaries of the lesion area to ensure that the tumor area in the image is fully and accurately represented. In order to ensure the consistency and reliability of the outline results, the consistency of the outline will be evaluated using the intraclass correlation coefficient (ICC). Only when the outline consistency meets the preset standard will the image data enter the subsequent feature extraction step.

[0041] After delineating the ROI, image analysis tools were used to extract radiomic features from breast ultrasound images. Radiomic features are derived by calculating and analyzing information such as the texture and shape of the lesion area in the image. They can provide a quantitative description of the biological characteristics of the tumor. To ensure the standardization and efficiency of the extraction process, the open source Python toolkit Pyradiomics was used in this study. This toolkit is widely used in medical image analysis and can automatically extract a large number of radiomic features from images. These features include, but are not limited to, the following categories: Shape characteristics: such as the major and minor axes, circumference, and area of ​​the tumor, reflecting the geometric shape of the tumor.

[0042] Gray-level co-occurrence matrix features: such as joint energy, maximum correlation coefficient, etc., reflect the spatial distribution of gray levels in the image and its texture characteristics.

[0043] Grayscale run length matrix feature: describes the regularity of the continuous distribution of grayscale values ​​in the image.

[0044] Grayscale size area matrix features: reflects the changes in grayscale value distribution in different areas of the image.

[0045] Neighborhood grayscale difference matrix features: describes the differences in grayscale values ​​and texture characteristics within a local area of ​​the image.

[0046] All extracted radiomics features are processed using Z-score normalization. Z-score normalization transforms the numerical value of each feature to conform to a standard normal distribution, preventing uneven impacts on subsequent analysis due to different dimensionality of different features. This processing step helps improve feature comparability and analytical accuracy.

[0047] The standardized radiomics features will provide data support for the subsequent LASSO regression model for feature screening and risk prediction. Through this step, this embodiment can extract radiomics features that can reflect the biological behavior and pathological characteristics of tumors, providing an effective quantitative basis for breast cancer recurrence risk assessment.

[0048] In step S3, the extracted radiomics features are first screened to select target feature variables that are highly correlated with breast cancer recurrence risk prediction. To achieve this, the LASSO regression (Least Absolute Shrinkage and Selection Operator) method is used in this example to screen the radiomics features.

[0049] LASSO regression is a regression method that effectively selects features. By introducing an L1 regularization term, it constrains the regression coefficients, shrinking irrelevant or redundant feature coefficients to zero while retaining features that significantly contribute to the prediction of the target variable (i.e., breast cancer recurrence risk). Radiomics data contains a large number of features, some of which may be redundant or unrelated to recurrence risk. Therefore, LASSO regression is used to screen these features to obtain the most predictive feature set.

[0050] Through LASSO regression, radiomic features are ranked according to their correlation with breast cancer recurrence risk, and a set of features with significant predictive value is selected. During this process, LASSO regression controls the degree of feature selection by adjusting the regularization parameter λ (lambda). Specifically, LASSO regression can automatically identify which radiomic features are associated with the risk of recurrence in breast cancer patients and assign non-zero regression coefficients to these features. At the same time, for features that are not significantly correlated with the risk of recurrence, LASSO regression shrinks their regression coefficients to zero, thereby eliminating these features.

[0051] After LASSO regression selects the target features, the size of the tumor is further determined by imaging features. Tumor size is an important parameter in the prognosis assessment of breast cancer, which can reflect the invasiveness and spread potential of the tumor. In this embodiment, the determination of tumor size will rely on the geometric features in the imaging features, especially the shape and size information of the tumor. This information is reflected by the shape features extracted from the imaging features (such as the major and minor axes, circumference, area, etc. of the tumor).

[0052] Preferably, the imaging omics features screened by LASSO regression in the present invention include: original_shape2D_MajorAxisLength (2D shape major axis length): The length of the major axis of the image ROI (region of interest) in the 2D plane, usually corresponding to the major axis length of the ellipse fit.

[0053] original_shape2D_PerimeterSurfaceRatio (two-dimensional shape perimeter-area ratio): The ratio of the perimeter of the ROI to its area (or surface area) in the 2D plane, reflecting the complexity of the outline relative to the internal area.

[0054] original_glcm_JointEnergy (Gray Level Co-occurrence Matrix - Joint Energy): Based on the "energy" feature defined by the Gray Level Co-occurrence Matrix (GLCM), also known as the Angular Second Moment. A larger value indicates a more concentrated grayscale distribution.

[0055] original_glcm_MCC (Gray Level Co-occurrence Matrix - Maximum Correlation Coefficient, Maximum Correlation Coefficient): An indicator that measures the maximum linear correlation between grayscales. A higher value indicates that the texture has stronger grayscale correlation.

[0056] original_glrlm_ShortRunLowGrayLevelEmphasis (Gray Level Run Length Matrix - Short Run Low Gray Level Emphasis): Based on the features of GLRLM (Gray Level Run Length Matrix), it is used to reflect the prominence of short-run low gray level runs in the image.

[0057] original_glszm_GrayLevelVariance (Gray Level Size Zone Matrix - Gray Level Variance): Based on the characteristics of GLSZM (Gray Level Size Zone Matrix), it measures the variance of all grayscale values ​​within each large cell, reflecting the degree of dispersion of the grayscale distribution.

[0058] original_glszm_LowGrayLevelZoneEmphasis (grayscale zone matrix - low grayscale zone emphasis): describes the frequency or prominence of low grayscale value areas in all zones.

[0059] original_glszm_SizeZoneNonuniformity (grayscale size zone matrix - region size nonuniformity): reflects the uniformity of the distribution of grayscale regions of different sizes. The higher the nonuniformity, the greater the difference in the distribution of grayscale regions of different sizes.

[0060] original_glszm_SmallAreaHighGrayLevelEmphasis (grayscale size area matrix - small area high grayscale emphasis): focuses on describing the significance of small area and high grayscale value areas.

[0061] original_glszm_ZoneEntropy (Grayscale Size Zone Matrix - Zone Entropy): Measures the randomness or uncertainty of the grayscale size zone distribution. A higher value indicates a more chaotic or complex texture.

[0062] original_glszm_ZonePercentage (grayscale zone matrix - zone percentage): measures the distribution of the effective grayscale area in the ROI in the proportion of all pixels in the image.

[0063] original_ngtdm_Busyness (Neighborhood Gray Tone Difference Matrix - Busyness): Based on the characteristics of NGTDM (Neighborhood Gray Tone Difference Matrix), a higher "busyness" indicates that the pixel grayscale changes more frequently within the local neighborhood.

[0064] original_ngtdm_Coarseness (neighborhood grayscale difference matrix - roughness): reflects the smoothness or roughness of the grayscale changes in the image within the local neighborhood. A larger value usually means a smoother texture.

[0065] original_ngtdm_Complexity (Neighborhood Grayscale Difference Matrix - Complexity): Measures the complexity of the grayscale difference distribution between neighbors. A larger value indicates more irregular local grayscale changes.

[0066] Based on the selected radiomic features, a radiomics scoring model can be constructed to quantify their contribution to the risk of breast cancer recurrence. The radiomics scoring model calculates each patient's radiomics score through a weighted summation of the features selected by LASSO regression and their regression coefficients. This scoring model comprehensively considers the impact of each radiomics feature to produce a comprehensive score that reflects the patient's tumor characteristics.

[0067] Preferably, the radiomics feature variables and coefficients screened above are shown in Table 1, and the radiomics score of the radiomics feature can be calculated by weighted summation.

[0068] Table 1 Variables and coefficients of the radiomics scoring formula In step S4, a nomogram prediction model was constructed based on the radiomics scores and clinicopathological parameters obtained in step S3. The core of this step is to combine the radiomics scores with clinicopathological parameters and construct a comprehensive risk prediction model through multivariate logistic regression analysis, further enabling accurate prediction of breast cancer recurrence risk.

[0069] First, clinicopathological parameters and radiomics scores were used as input variables in a multivariate logistic regression model for analysis. Specifically, clinicopathological parameters included patient age, number of lymph node metastases, ER status, PR status, Her-2 status, Ki67 index, histological grade, and vascular invasion, reflecting the patient's clinical background and biological characteristics. The radiomics score, calculated by weighted summation of the aforementioned radiomics features, provides a quantitative description of the patient's tumor characteristics. The combination of the two can comprehensively assess the risk of recurrence in breast cancer patients.

[0070] When performing multivariate logistic regression, the model evaluates the contribution of each characteristic variable to the risk of breast cancer recurrence. Through regression analysis, the model assigns corresponding regression coefficients to each characteristic, which reflect the importance and direction (positive or negative impact) of the characteristic in predicting the risk of breast cancer recurrence. In this way, the model can calculate the patient's total score based on clinical parameters and imaging data.

[0071] Next, the regression coefficients obtained from the regression analysis were multiplied by the actual values ​​of each variable, and the sum was finally calculated to obtain a composite score. This score represents the patient's overall risk of recurrence and serves as a key input in the nomogram model, determining whether the patient falls into the high-risk or low-risk category.

[0072] The nomogram model constructed in the present invention is as follows Figure 2 As shown in Figure 2, the nomogram model maps the contribution of each variable to a corresponding quantitative score, creating an intuitive and easy-to-use tool. By inputting a patient's clinical parameters and radiomics score, physicians can quickly calculate the patient's overall score and, based on this score, identify the corresponding risk category on the graph. This classification can inform clinical decision-making, helping physicians develop personalized treatment plans and ultimately achieve precision medicine.

[0073] When constructing the nomogram model, the scoring system for all variables underwent rigorous statistical validation to ensure its high reliability and accuracy. During the multivariate logistic regression analysis, appropriate statistical methods were used to test the association between each variable and the risk of breast cancer recurrence. Variables were processed using statistical methods such as the chi-square test, t-test, or Mann-Whitney U test to ensure statistical significance of the model.

[0074] In summary, step S4 constructed a nomogram prediction model by combining radiomics scores with clinicopathological parameters using multivariate logistic regression analysis. This model transforms complex clinical information and imaging data into a simple and practical tool, providing a scientific and reliable basis for recurrence risk assessment in breast cancer patients and helping physicians make more accurate treatment decisions.

[0075] In step S5, based on the nomogram model constructed in step S4 and the calculated total score (TotalPoint), the patient's recurrence risk probability value is further calculated by a linear predictor (Linear Predictor), and the patient is divided into a high-risk category or a low-risk category accordingly.

[0076] First, each input feature variable is mapped to the nomogram scoring system. Based on the pre-defined scoring criteria in the nomogram model, the score corresponding to each feature is read from the "Points" axis. Input features include radiomics score (radiomics) and tumor size (size), which are derived from radiomics analysis results and are used to characterize the internal texture and geometric characteristics of the tumor. They also include multiple clinicopathological parameters, such as patient age (age), number of lymph node metastases (nLN), estrogen receptor status (ER), progesterone receptor status (PR), human epidermal growth factor receptor 2 status (Her-2), proliferation index (Ki67), histological grade (histo_grade), and vascular invasion (lymph_vessel_invasion). Each of these features is assigned a corresponding scoring scale in the nomogram model to quantify its weight in risk prediction.

[0077] After mapping each characteristic parameter, its corresponding quantitative score is read one by one, and the scores of all characteristic parameters are accumulated to obtain a total score (Total Point) representing the patient's overall risk level. This total score serves as input to the nomogram model, reflecting the patient's position in the multidimensional characteristic space and is used to further determine the risk level.

[0078] This total score is then calculated using a pretrained linear predictor to estimate the patient's risk of distant recurrence. The linear predictor's core mechanism involves inputting the total score into a linear function composed of the model coefficients determined during the multivariate logistic regression training phase. This function is a linear combination of a set of fixed parameters and the score. This function then outputs a risk probability value between 0 and 1, representing the patient's likelihood of developing distant recurrence within a specific time period (e.g., 10 years).

[0079] Furthermore, the outputted risk probability value is compared with the risk classification threshold to determine the risk category to which the patient belongs. Preferably, when the predicted probability is higher than the high-risk threshold preset by the model, the patient is classified as a high-risk category, which means that he or she has a higher risk of distant recurrence in the next few years (such as 10 years); otherwise, he or she is classified as a low-risk category, indicating that the possibility of distant recurrence is low. The division of risk categories is of great significance in clinical applications and can provide data support for the patient's subsequent treatment decisions, monitoring strategies and individualized management.

[0080] In the risk classification process, the probability mapping relationship based on the linear predictor is derived from the statistical fitting results in the previous nomogram modeling process. This result combines the actual follow-up status and characteristic performance of the samples in the training data set and has certain representativeness and generalizability.

[0081] In summary, in step S5, the multi-parameter scoring results obtained in step S4 are input into the linear predictor of the nomogram model to calculate the patient's risk probability value, and high- and low-risk stratification is performed accordingly. This method integrates complex, multidimensional structured and unstructured medical information and transforms it into an intuitive clinical risk assessment indicator, providing a practical and feasible technical path for assisting in the determination of breast cancer recurrence risk.

[0082] Example: To better illustrate the technical solution of the present invention, the following provides an embodiment of the present invention in combination with a specific experimental process. This embodiment combines radiomics features with clinical pathological parameters to construct a nomogram model for predicting the risk of distant recurrence of breast cancer, and trains and validates the model.

[0083] 1. Selection of experimental patients In this example, the experimental subjects are hormone receptor-positive / human epidermal growth factor receptor 2-negative (HR+ / HER2-) early breast cancer patients who undergo MammaPrint testing. The selected patients must meet the following conditions: Preferably, the patient's Adjuvant Online score is high, indicating a potential risk of recurrence; Patients did not receive any intervention procedures such as neoadjuvant chemotherapy, radiotherapy, or puncture biopsy before breast ultrasound examination to exclude the influence of operational factors on tumor imaging characteristics; The lesion region (ROI) in the ultrasound image is completely visible, and the image quality meets the analysis requirements; The clinical pathological data are complete, which facilitates subsequent statistical modeling and feature integration.

[0084] 2. Determine radiomics features 1) High-performance ultrasound equipment, such as the Samsung WS80A, Philips EPIQ 5 & 7, Philips IU22, and GE Logic9, equipped with a high-frequency linear array probe (frequency range 3-12 MHz) with a preferred center frequency of 10 MHz, was used. Examinations were performed by professional sonographers who have undergone systematic breast ultrasound training to ensure consistent image quality. To ensure the integrity of the radiomics analysis area, patients with tumors larger than 5 cm were excluded to prevent the possibility that a single ultrasound frame would not fully capture the lesion area.

[0085] 2) Using ImageJ, an open-source imaging software developed by the National Institutes of Health (NIH), lesion regions were manually delineated and relevant features were extracted. To verify the consistency of the delineation, the intraclass correlation coefficient (ICC) was used to assess the consistency, and only features with an ICC greater than 0.75 were retained to ensure the stability of feature extraction.

[0086] 3) Using the open-source Python toolkit Pyradiomics, a total of 105 radiomic features were extracted from the delineated ROIs. These features included shape features, gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighborhood gray-level difference matrix (NGTDM). After extraction, all feature variables were normalized using the z-score method to eliminate the influence of different dimensions.

[0087] 4) All patient samples were randomly divided into training and test sets in a 7:3 ratio. Within the training set, 105 radiomic features were screened using the LASSO regression model. Multicollinearity between variables was assessed using the variance inflation factor (VIF). For variables with collinearity, only one was retained to avoid redundancy affecting model stability.

[0088] 5) The following 14 radiomics feature variables with potential predictive value and their coefficients were retained through LASSO screening, as shown in Table 2: Table 2 Variables, coefficients and Spearman rank correlation coefficients of the radiomics scoring formula The coefficients of these 14 characteristic variables were determined through a regression model and used to construct a radiomics scoring formula. Each patient's radiomics score was calculated by multiplying the above characteristic variables by their corresponding coefficients and summing them. This score was then combined with the maximum tumor diameter, a tumor geometry indicator, to form a complete radiomics feature set.

[0089] 3. Determination of clinicopathological parameters In terms of clinical parameters, eight representative indicators were selected: age, number of metastatic lymph nodes (nLN), estrogen receptor status (ER), progesterone receptor status (PR), Her-2 status, Ki67 proliferation index, histological grade (histo_grade), and vascular invasion (lymph_vessel_invasion). These indicators provide basic information for breast cancer classification and prognosis.

[0090] 4. Construct and validate the nomogram model 1) If Figure 3 、 4 As shown in Figures 5 and 5 , based on the above characteristic variables, three nomogram models were constructed for comparative analysis: one was a radiomics model, with the input variables being radiomics score and tumor size, which was used to evaluate the ability of imaging dimension data to predict risk; The second is a clinical model, which uses all eight clinical pathological parameters as input variables to evaluate the predictive performance of traditional clinical data; the third is an imaging-clinical joint model, which integrates all variables of the first two to construct a comprehensive risk assessment model.

[0091] The nomogram model was constructed using R language combined with EmpowerStats software. During model development, categorical variables were analyzed for intergroup differences using the chi-square test, while continuous variables were compared using either the t-test or the Mann-Whitney U test based on their distribution characteristics. All statistical analyses were performed using a two-sided approach, with a preferred significance level of p < 0.05.

[0092] 2) To verify the performance of the nomogram prediction model constructed in the present invention, a systematic evaluation was performed on three models in the training set and validation set, including: an imaging-clinical joint model, a clinical model, and an imaging omics model.

[0093] During the validation process, the receiver operating characteristic (ROC) curve and its area under the curve (AUC) were used to comprehensively evaluate the ability of each model to discriminate the risk of distant breast cancer recurrence. Furthermore, the model classification performance was quantitatively analyzed using metrics such as sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy.

[0094] like Figure 3-Figure 9 The results showed that the combined imaging-clinical model achieved an AUC of 0.899 (95% confidence interval, 0.8451–0.9523) in the training set and 0.846 (95% confidence interval, 0.7362–0.9561) in the validation set, outperforming the other two models overall. Specifically, the model achieved a sensitivity of 72.7%, a specificity of 92.9%, a positive predictive value of 88.9%, a negative predictive value of 81.3%, and an overall accuracy of 84.0% in the training set. In the validation set, its sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were 85.2%, 76.9%, 79.3%, 83.3%, and 81.1%, respectively.

[0095] In contrast, the clinical model had an AUC of 0.845 in the training set and 0.806 in the validation set, with corresponding accuracies of 76.8% and 75.5%, respectively. The imaging omics model had an AUC of 0.800 (training set) and 0.662 (validation set), respectively, showing relatively low predictive performance in the validation set, with an accuracy of 66.0%.

[0096] In addition, a statistical test was performed to examine the differences in AUC between the training and validation sets for the three models. The results showed that there was no statistically significant difference in AUC between the two datasets for the three models (imaging-clinical model: p = 0.402; clinical model: p = 0.583; radiomics model: p = 0.11), indicating that the models exhibited good stability and generalization capabilities in the validation set. The evaluation metrics for each model on the training and validation sets are shown in Table 3: Table 3 Evaluation indicators of each model on the training set and validation set Among them, ALL is the imaging-clinical model, CLINIC is the clinical model, RADIOMICS is the imaging omics model; Tra is the training set, and Val is the validation set.

[0097] Furthermore, to evaluate the clinical utility of the model at different cutoff probabilities (i.e., high-risk thresholds), precision-recall (PR) curves were plotted for each model. The PR curves, in both the training and validation sets, showed that the combined imaging-clinical model performed better in terms of overall balance between recall and precision, making it suitable for high-risk patient identification scenarios.

[0098] Model calibration analysis used a bootstrap method (with 500 replicates) to generate a calibration curve. This included comparisons of the apparent calibration line, the bias-corrected line, and the ideal line to assess the consistency between the model's predicted probabilities and actual observed outcomes. The results showed good consistency between the predicted probabilities and actual risks of the combined imaging-clinical model in both the training and validation sets, demonstrating good calibration performance.

[0099] Furthermore, to evaluate the model's applicability in real-world clinical scenarios, a decision curve analysis and standardized net benefit curve were constructed. This analysis quantified the model's net benefit based on different high-risk cutoff probabilities and the corresponding false-positive cost-to-true-positive benefit ratios. Results from both the training and validation sets demonstrated that the combined imaging-clinical model demonstrated a high net benefit at most thresholds.

[0100] Finally, the model's clinical impact was further analyzed using a clinical impact curve. In this analysis, the vertical axis represents the number of patients predicted by the model as high-risk per 1,000 patients, the red line represents the ideal test (i.e., the reference MammaPrint test results), and the black line represents the model's actual predictions. The results showed that the combined imaging-clinical model closely aligned with the ideal line, with a small uncertainty interval, suggesting that the model has high potential for clinical application.

[0101] In summary, the combined imaging-clinical model proposed in this paper demonstrates superior performance compared to traditional clinical models and simple imaging omics models in discriminating the risk of distant recurrence of breast cancer. It has good stability, calibration, and clinical practicality, and is suitable for assisting in the formulation of clinical individualized treatment plans.

[0102] like Figure 10 As shown, the present invention also provides a breast cancer risk assessment system, comprising: An image acquisition module 10 is used to acquire breast ultrasound images; Pathology information acquisition module 20, used to obtain clinical pathology parameters of patients; A feature extraction module 30 extracts radiomic features based on the ROI region of the breast image; Risk modeling module 40, which uses radiomics scores and clinical parameters to construct a nomogram model; The risk output module 50 outputs a high / low risk assessment result based on the total score.

[0103] The system of this embodiment can be used to execute the above method embodiments, and its principles and technical effects are similar, so they will not be repeated here.

[0104] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A 70-gene breast cancer risk assessment method based on imaging and pathology, characterized by: The following steps are involved: 1) Obtain clinicopathological parameters of patients with HR+ / HER2- early breast cancer; 2) Obtain the patient's breast ultrasound image, delineate the lesion area, and extract radiomics features; 3) LASSO regression was used to select target feature variables from the extracted radiomics features, tumor size was determined based on the radiomics features, and a radiomics score was constructed; 4) Combining clinical pathological parameters with radiomics scores, a nomogram prediction model was constructed through multivariate logistic regression, and each parameter was mapped into a quantitative score to obtain a total score; 5) Based on the total score, the risk probability value is calculated using a linear predictor to classify the patients into high-risk or low-risk categories.

2. The method for risk assessment of breast cancer 70 genes based on imaging and pathology according to claim 1, characterized in that: The clinical pathological parameters include: Patient age, number of lymph node metastases, ER status, PR status, Her-2 status, Ki67 index, histological grade, and vascular invasion; The clinical pathological parameters were collected before the patients underwent breast ultrasound examination, and the patients did not receive neoadjuvant chemotherapy, radiotherapy, puncture biopsy or local breast surgery.

3. The method for risk assessment of breast cancer 70 genes based on imaging and pathology according to claim 1, characterized in that: The breast ultrasound images are collected by medical ultrasound equipment with high-resolution imaging capabilities, and the lesion images are delineated by imaging professionals for ROI. The consistency of the ROI delineation is evaluated by the intra-group correlation coefficient. Only when the intra-group correlation coefficient is higher than a preset consistency threshold, the image data enters the subsequent feature extraction step.

4. The method for risk assessment of breast cancer 70 genes based on imaging and pathology according to claim 1, characterized in that: The radiomics features are extracted by image analysis tools, and the radiomics features include: Shape features, gray-level co-occurrence matrix features, gray-level run length matrix features, gray-level size area matrix features, neighborhood gray-level difference matrix features; After the imaging omics features were standardized by Z-score, LASSO regression was used for variable screening, and the variance inflation factor was used to eliminate collinear variables, and the features that met the conditions were retained for model construction.

5. The method for risk assessment of breast cancer 70 genes based on imaging and pathology according to claim 1, characterized in that: The radiomics score is obtained by a radiomics scoring model, which is constructed based on characteristic variables screened by LASSO regression, including: 2D shape main axis length, 2D shape perimeter-area ratio, gray-level co-occurrence matrix joint energy, gray-level co-occurrence matrix maximum correlation coefficient, gray-level run length matrix short-range low gray-level emphasis, gray-level size area matrix gray-level variance, gray-level size area matrix low gray-level area emphasis, gray-level size area matrix regional size inhomogeneity, gray-level size area matrix small area high gray-level emphasis, gray-level size area matrix regional entropy, gray-level size area matrix regional percentage, neighborhood gray-level difference matrix busyness, neighborhood gray-level difference matrix roughness, neighborhood gray-level difference matrix complexity; The score is obtained by multiplying the value of each variable by its LASSO regression coefficient and then summing the weighted values.

6. The method for risk assessment of breast cancer 70 genes based on imaging and pathology according to claim 1, characterized in that: The correlation between the target characteristic variables screened out in the imaging omics features and the risk classification results of the 70-gene test was verified by Spearman rank correlation coefficient analysis and was statistically significant.

7. The method for risk assessment of breast cancer 70 genes based on imaging and pathology according to claim 1, characterized in that: The nomogram prediction model is constructed using a multivariate logistic regression method, comprising the following steps: 1) Input candidate predictor variables into the logistic regression model and perform multivariate stepwise regression analysis; 2) converting modeling variables into a quantitative scoring system and accumulating the patient's total score; 3) The total score is converted into a predicted probability value based on the linear prediction function, and a nomogram is drawn to achieve individualized risk prediction.

8. The method for risk assessment of breast cancer 70 genes based on imaging and pathology according to claim 1, characterized in that: The nomogram model was constructed in the following three versions for comparative verification: Radiomics models constructed based only on radiomics scores and tumor size; Clinical models constructed based solely on clinicopathological parameters; A joint imaging-clinical model constructed by integrating imaging scores and clinical parameters; The imaging omics model, clinical model and imaging-clinical joint model are all constructed based on the same sample data set.

9. The method for risk assessment of breast cancer 70 genes based on imaging and pathology according to claim 1, characterized in that: The risk probability value output by the model is a continuous value between 0 and 1; Based on empirical judgment criteria, when the probability value is higher than the median value, the patient is judged to be in the high-risk category of the 70-gene breast cancer test and is suitable for subsequent intensive treatment; When the probability value is lower than the median value, it is judged as a low-risk category, and unnecessary chemotherapy intervention can be avoided.

10. A breast cancer risk assessment system for implementing the method according to any one of claims 1 to 9, characterized in that: include: An image acquisition module, used for acquiring breast ultrasound images; Pathology information acquisition module, used to obtain the patient's clinical pathology parameters; Feature extraction module, which extracts radiomic features based on the ROI region of the breast image; The risk modeling module uses radiomics scores and clinical parameters to construct a nomogram model; The risk output module outputs high / low risk assessment results based on the total score.

Images