Method for predicting prognosis degree of MDA5 + DM-ILD by using PET / CT image score

Abstract

The invention discloses a standardized method for predicting the prognosis degree of MDA5 + DM-ILD by using PET / CT image scoring. The method comprises the following steps: an image data acquisition and extraction step: acquiring PET / CT-based image data of a sample; the PET / CT-based image data are a plurality of lung image data; the method comprises the following steps: sketching the contour of a focus layer by layer through image segmentation and drawing, then performing three-dimensional volume recombination on a two-dimensional focus region to generate a three-dimensional interested volume, and extracting feature data from the three-dimensional interested volume; an image data analysis step: classifying the obtained feature data according to radiomics features, importing the classified feature data into a prediction model, improving model prediction by using a Pearson correlation coefficient through an improved random forest algorithm, and outputting scoring index data of a sample according to deep learning; and performing cooperative verification according to a training result and biochemical indexes, blood routine indexes and marker concentration data of the sample. Through the datamation analysis method provided by the invention, the medical cost can be effectively reduced and the medical efficiency can be effectively improved through datamation verification.

Description

Technical Field

[0001] This invention relates to the field of PET / CT image analysis and calculation, and specifically to a method for predicting the prognostic severity of MDA5+DM-ILD using PET / CT image scoring. Background Technology

[0002] Anti-MDA5 (melanin differentiation-associated gene) positive dermatomyositis is an autoimmune disease characterized by muscle weakness and a distinctive rash, while interstitial lung disease is a common and serious complication of this disease. The presence of anti-MDA5 antibodies is associated with a more aggressive and poorer prognostic form of interstitial lung disease in these patients, which may lead to a poorer outcome if not properly managed.

[0003] Radiomics, also known as radiomics, refers to the extraction and analysis of large amounts of quantitative imaging features using methods such as computed tomography (CT), positron emission tomography (PET), or magnetic resonance imaging (MRI) to obtain high-throughput imaging feature data for medical images. It involves transforming imaging data into analyzable digital features and then analyzing them using methods from computer science, statistics, and machine learning. The core of radiomics lies in extracting large amounts of quantitative data from conventional images (such as CT, MRI, or PET scans) that reflect the morphology, texture, and other imaging characteristics of tissues or organs, thereby revealing potential disease information.

[0004] With the application of machine learning technology in the analysis of interstitial lung disease, the accuracy of interstitial lung disease analysis has been further improved.

[0005] In current technology, the primary method for examining interstitial lung disease (ILD) is chest CT scan, especially high-resolution CT (HRCT) scans. CT scans can reveal structural changes in the lungs, such as inflammation, fibrosis, and alveolar wall thickening. In the early stages of ILD, imaging findings may be atypical or overlap with other diseases (such as pneumonia or pulmonary edema). Furthermore, different doctors and hospitals may have different standards for interpreting images, requiring standardized training and experience. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, the purpose of this invention is to provide a standardized method for predicting the prognosis of MDA5+DM-ILD using PET / CT image scoring.

[0007] The aim is to assess the prognosis of MDA5+DM-ILD (dermatomyositis-interstitial lung disease) by using standardized data processing and deep learning algorithms to quantify CT image features.

[0008] A standardized method for predicting the prognostic severity of MDA5+DM-ILD using PET / CT image scoring includes: Image data acquisition and extraction steps: Acquire PET / CT-based image data for the sample; The PET / CT-based imaging data consists of several lung imaging data. Lung imaging data includes density, morphology, and / or texture imaging data of different types of lung lesions (ground-glass opacity, solid opacity, fibrosis, etc.); By segmenting and drawing images, the outline of the lesion is delineated layer by layer. Then, the two-dimensional lesion area is reconstructed into a three-dimensional volume to generate a three-dimensional volume of interest, and feature data is extracted from it. Image data analysis steps: The acquired feature data is classified according to radiomics features and then imported into the prediction model. The improved random forest algorithm is used to improve the model prediction by utilizing the Pearson correlation coefficient. The sample score index data is output based on deep learning. The training results were validated in conjunction with the biochemical indicators, blood routine indicators, and biomarker concentration data of the samples.

[0009] In a preferred embodiment of the present invention, the different types of lung lesions are ground-glass opacities and / or solid lesions and / or fibrotic lesions.

[0010] In a preferred embodiment of the present invention, the feature data includes: Histogram intensity of lesion area; Shape of the histogram of the lesion area; Histogram size characteristics of the lesion area; The texture and structure of the lesion area are either homogeneous or heterogeneous; Localization of the lesion area; Fractal of the lesion area; Characteristics of lesion area fusion; The relationship between the lesion area and the surrounding tissues; Is the lesion area adjacent to the bronchial tree and vascular tree of the lungs?

[0011] In a preferred embodiment of the present invention, the improved random forest algorithm is to construct an optimal decision tree ensemble by combining random forest with a genetic algorithm; The construction of each tree in the random forest repeats the following steps: (1) Perform a bootstrap resampling on the training set to obtain a sub-training set; (2) Use the sub-training set obtained in (1) to start training a decision tree; (3) At each node of the decision tree, randomly select N features and select the best one as the current node by calculating their information gain; (4) Repeat (3) until a tree is built; The ensemble error rate of this model is: training of the model ends when the error rate is less than 0.1. After the model training is complete, the training results are output.

[0012] In a preferred embodiment of the present invention, the step of randomly selecting N features and selecting the optimal one as the current node by calculating their information gain is as follows: N is the integer part of the square root of all feature numbers p, i.e., N = round(sqrt(p)), and N features are selected as candidate features from all features using uniform random sampling without replacement.

[0013] In a preferred embodiment of the present invention, the fit verification step includes: The prognostic risk index output by the predictive model was compared and validated with the patient's immunological and inflammatory biochemical indicators to determine whether the model output results were consistent with clinical prognostic indicators.

[0014] The preferred inflammatory biochemical indicators are antibody titer MDA5 or ferritin.

[0015] In a preferred embodiment of the present invention, the step of verifying the training results in conjunction with the biochemical indicators, blood routine indicators, and biomarker concentration data of the samples specifically involves: During model training, when any of the following conditions are met, it is necessary to further collect and incorporate biochemical indicators, routine blood indicators, and disease-related biomarker concentration data from the samples to jointly validate and calibrate the model output: Condition one: When there is a discrepancy between the image score index output by the model and the actual follow-up outcome, When the high-risk score predicted by the model is inconsistent with the actual progression of the patient's condition, it is necessary to introduce biochemical indicators to assess the influencing factors. Condition two: When the ranking of feature importance shifts or the model stability decreases, the correlation with biochemical / blood routine indicators is compared to determine whether the model has missed key clinical features.

[0016] Condition three: When the model is at risk of overfitting, biochemical and biomarker data are used as an independent external validation set to test whether the image features have reliable generalization ability.

[0017] A standardized management system for predicting the prognostic severity of MDA5+DM-ILD using PET / CT image scoring includes: Image data acquisition and extraction module: Acquire PET / CT-based image data for the sample; The PET / CT-based imaging data consists of several lung imaging data. Lung imaging data includes density, morphology, and / or texture imaging data of different types of lung lesions (ground-glass opacity, solid opacity, fibrosis, etc.); By segmenting and drawing images, the outline of the lesion is delineated layer by layer. Then, the two-dimensional lesion area is reconstructed into a three-dimensional volume to generate a three-dimensional volume of interest, and feature data is extracted from it. Image data analysis module: The acquired feature data is classified according to radiomics features and then imported into the prediction model. The improved random forest algorithm is used to improve the model prediction by utilizing the Pearson correlation coefficient. The sample score index data is output based on deep learning. The training results were validated in conjunction with the biochemical indicators, blood routine indicators, and biomarker concentration data of the samples.

[0018] In a preferred embodiment of the present invention, the different types of lung lesions are ground-glass opacities and / or solid lesions and / or fibrotic lesions.

[0019] In a preferred embodiment of the present invention, the feature data includes: Histogram intensity of lesion area; Shape of the histogram of the lesion area; Histogram size characteristics of the lesion area; The texture and structure of the lesion area are either homogeneous or heterogeneous; Localization of the lesion area; Fractal of the lesion area; Characteristics of lesion area fusion; The relationship between the lesion area and the surrounding tissues; Is the lesion area adjacent to the bronchial tree and vascular tree of the lungs?

[0020] In a preferred embodiment of the present invention, the improved random forest algorithm is to construct an optimal decision tree ensemble by combining random forest with a genetic algorithm; The construction of each tree in the random forest repeats the following steps: (1) Perform a bootstrap resampling on the training set to obtain a sub-training set; (2) Use the sub-training set obtained in (1) to start training a decision tree; (3) At each node of the decision tree, randomly select N features and select the best one as the current node by calculating their information gain; (4) Repeat (3) until a tree is built; The ensemble error rate of this model is: training of the model ends when the error rate is less than 0.1. After the model training is complete, the training results are output.

[0021] In a preferred embodiment of the present invention, the step of randomly selecting N features and selecting the optimal one as the current node by calculating their information gain is as follows: N is the integer part of the square root of all feature numbers p, i.e., N = round(sqrt(p)), and N features are selected as candidate features from all features using uniform random sampling without replacement.

[0022] In a preferred embodiment of the present invention, the fit verification step includes: The prognostic risk index output by the predictive model was compared and validated with the patient's immunological and inflammatory biochemical indicators to determine whether the model output results were consistent with clinical prognostic indicators.

[0023] The preferred inflammatory biochemical indicators are antibody titer MDA5 or ferritin.

[0024] In a preferred embodiment of the present invention, the step of verifying the training results in conjunction with the biochemical indicators, blood routine indicators, and biomarker concentration data of the samples specifically involves: During model training, when any of the following conditions are met, it is necessary to further collect and incorporate biochemical indicators, routine blood indicators, and disease-related biomarker concentration data from the samples to jointly validate and calibrate the model output: Condition one: When there is a discrepancy between the image score index output by the model and the actual follow-up outcome, When the high-risk score predicted by the model is inconsistent with the actual progression of the patient's condition, it is necessary to introduce biochemical indicators to assess the influencing factors. Condition two: When the ranking of feature importance shifts or the model stability decreases, the correlation with biochemical / blood routine indicators is compared to determine whether the model has missed key clinical features.

[0025] Condition three: When the model is at risk of overfitting, biochemical and biomarker data are used as an independent external validation set to test whether the image features have reliable generalization ability.

[0026] The beneficial effects of this invention are as follows: The data analysis method of this invention can effectively help reduce medical costs and improve medical efficiency through data verification. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. Furthermore, in the following descriptions, well-known structures and technologies are omitted to avoid unnecessarily obscuring the concept of the invention.

[0029] By acquiring PET / CT images of patients with interstitial lung disease, an improved random forest algorithm was established. After the algorithm was established, PET / CT analysis was used to classify the type (ground-glass opacity, solid, fibrosis) and degree (proportion of the total lesion area). This method can improve the efficiency of prognostic assessment and avoids unnecessary examinations for patients.

[0030] Based on PET / CT image data, image acquisition and reconstruction, image segmentation and rendering, feature extraction and quantization, database establishment and data sharing are performed. An improved random forest algorithm is established, and Pearson correlation coefficient is used to improve model prediction. A scoring index is calculated based on deep learning to improve the efficiency of prognostic assessment.

[0031] Specifically as follows: A standardized method for predicting the prognostic severity of MDA5+DM-ILD using PET / CT image scoring includes: Image data acquisition and extraction steps: Acquire PET / CT-based image data for the sample; The PET / CT-based imaging data consists of several lung imaging data. Lung imaging data includes density, morphology, and / or texture imaging data of different types of lung lesions (ground-glass opacity, solid opacity, fibrosis, etc.); The different types of lung lesions are ground-glass opacities and / or solid lesions.

[0032] By segmenting and drawing images, the outline of the lesion is delineated layer by layer. Then, the two-dimensional lesion area is reconstructed into a three-dimensional volume to generate a three-dimensional volume of interest, and feature data is extracted from it. Feature data includes: Histogram intensity of lesion area; Shape of the histogram of the lesion area; Histogram size characteristics of the lesion area; The texture and structure of the lesion area are either homogeneous or heterogeneous; Localization of the lesion area; Fractal of the lesion area; Characteristics of lesion area fusion; The relationship between the lesion area and the surrounding tissues; Is the lesion area adjacent to the bronchial tree and vascular tree of the lungs?

[0033] Image data analysis steps: The acquired feature data is classified according to radiomics features and then imported into the prediction model. The improved random forest algorithm is used to improve the model prediction by utilizing the Pearson correlation coefficient. The sample score index data is output based on deep learning. The improved random forest algorithm utilizes random forests combined with genetic algorithms to construct the optimal decision tree ensemble. The construction of each tree in the random forest repeats the following steps: (1) Perform a bootstrap resampling on the training set to obtain a sub-training set; (2) Use the sub-training set obtained in (1) to start training a decision tree; (3) At each node of the decision tree, randomly select N features and select the best one as the current node by calculating their information gain.

[0034] N is the integer part of the square root of all feature numbers p, i.e., N = round(sqrt(p)), and N features are selected as candidate features from all features using uniform random sampling without replacement.

[0035] (4) Repeat (3) until a tree is built; The ensemble error rate of this model is: training of the model ends when the error rate is less than 0.1. After the model training is complete, the training results are output.

[0036] The training results were validated in conjunction with the biochemical indicators, blood routine indicators, and biomarker concentration data of the samples.

[0037] When any of the following conditions are met, further data on biochemical indicators, routine blood indicators, and disease-related biomarker concentrations of the samples should be collected and incorporated to jointly validate and calibrate the model output: Condition one: When there is a discrepancy between the image score index output by the model and the actual follow-up outcome, When the high-risk score predicted by the model is inconsistent with the actual progression of the patient's condition, it is necessary to introduce biochemical indicators to assess the influencing factors. Condition two: When the ranking of feature importance shifts or the model stability decreases, the correlation with biochemical / blood routine indicators is compared to determine whether the model has missed key clinical features.

[0038] Condition three: When the model is at risk of overfitting, biochemical and biomarker data are used as an independent external validation set to test whether the image features have reliable generalization ability.

[0039] The verification process includes: The prognostic risk index output by the predictive model was compared and validated with the patient's immunological and inflammatory biochemical indicators to determine whether the model output results were consistent with clinical prognostic indicators.

[0040] The following is an illustration through specific examples: 1. Achieve time savings In a top-tier hospital, PET / CT images of 15 patients with MDA5+DM-ILD were analyzed.

[0041] The methods required by existing technologies are as follows: Radiologists need to manually delineate the extent of ILD lesions layer by layer; For a single CT image with 400-600 slices, manual delineation takes an average of 40-60 minutes per case. The 3D reconstruction requires approximately 10 minutes of manual inspection and correction.

[0042] Total time: approximately 50-70 minutes per case.

[0043] The method of this invention is as follows: An image segmentation algorithm based on PET / CT grayscale thresholding and texture consistency is used to automatically generate preliminary lesion regions; Random forest model for screening key radiomics features; The system automatically completes three-dimensional volume reconstruction; Manual fine-tuning of the final boundaries takes 3 to 5 minutes.

[0044] Total time: approximately 6-10 minutes per case.

[0045] Compared with the prior art, the advantages of the present invention are as follows: Early identification of high-risk patients and timely adjustment of treatment plans can significantly reduce the following costs: 1. ICU hospitalization costs (often 20,000–50,000 RMB per day) 2. Indiscriminate use of expensive drugs such as high-dose hormones, immunosuppressants, and targeted therapies. 3. Costs of life-saving equipment such as ventilators and ECMO. 4. Repeated examinations and hospitalization turnover costs The cost of a single PET / CT scan is 6,500 yuan, which is extremely low compared to the cost of an ICU treatment.

[0046] Moreover, in many cases: All patients were treated under a "high-risk" protocol (which was extremely expensive), even though some patients did not require such aggressive treatment, resulting in a large amount of "overtreatment costs".

[0047] This invention uses data analysis methods to accurately distinguish between high-risk and low-risk patients, effectively helping to reduce medical costs and improve medical efficiency.

[0048] High-risk individuals can have their treatment plans upgraded promptly, while low-risk individuals can avoid overuse of immunosuppressants and antifibrotic drugs. This "precise stratified management" reduces unnecessary expenditures, thereby lowering overall costs.

[0049] The foregoing has shown and described the basic principles and main features of the invention and the advantages of the invention.

[0050] Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope. All such changes and modifications fall within the scope of the present invention as claimed, which is defined by the appended claims and their equivalents.

Claims

1. A standardized method for predicting the prognostic severity of MDA5+DM-ILD using PET / CT image scoring, characterized in that, include: Image data acquisition and extraction steps: Acquire PET / CT-based image data for the sample; The PET / CT-based imaging data consists of several lung imaging data. Lung imaging data includes density, morphology, and / or texture imaging data of different types of lung lesions; By segmenting and drawing images, the outline of the lesion is delineated layer by layer. Then, the two-dimensional lesion area is reconstructed into a three-dimensional volume to generate a three-dimensional volume of interest, and feature data is extracted from it. Image data analysis steps: The acquired feature data is classified according to radiomics features and then imported into the prediction model. The improved random forest algorithm is used to improve the model prediction by utilizing the Pearson correlation coefficient. The sample score index data is output based on deep learning. The training results were validated in conjunction with the biochemical indicators, blood routine indicators, and biomarker concentration data of the samples.

2. The standardized method for predicting the prognostic severity of MDA5+DM-ILD using PET / CT image scoring as described in claim 1, characterized in that, The different types of lung lesions are ground-glass opacities and / or solid lesions and / or fibrotic lesions.

3. The standardized method for predicting the prognostic severity of MDA5+DM-ILD using PET / CT image scoring as described in claim 1, characterized in that, The feature data includes: Histogram intensity of lesion area; Shape of the histogram of the lesion area; Histogram size characteristics of the lesion area; The texture and structure of the lesion area are either homogeneous or heterogeneous; Localization of the lesion area; Fractal of the lesion area; Characteristics of lesion area fusion; The relationship between the lesion area and the surrounding tissues; Is the lesion area adjacent to the bronchial tree and vascular tree of the lungs? 4. A standardized method for predicting the prognostic severity of MDA5+DM-ILD using PET / CT image scoring as described in claim 1, characterized in that, The improved random forest algorithm utilizes random forest and combines it with a genetic algorithm to construct an optimal decision tree ensemble. The construction of each tree in the random forest repeats the following steps: (1) Perform a bootstrap resampling on the training set to obtain a sub-training set; (2) Use the sub-training set obtained in (1) to start training a decision tree; (3) At each node of the decision tree, randomly select N features and select the best one as the current node by calculating their information gain; (4) Repeat (3) until a tree is built; The ensemble error rate of this model is: training of the model ends when the error rate is less than 0.

1. After the model training is complete, the training results are output.

5. A standardized method for predicting the prognostic severity of MDA5+DM-ILD using PET / CT image scoring as described in claim 1, characterized in that, The aforementioned verification is to obtain the gene type for determining lung cancer.

Images