Image processing method and device based on multi-mode CMR imageomics and application

Abstract

According to the image processing method and device based on the multi-modal CMR imageomics and the application, accurate distinguishing of cardiovascular magnetic resonance images can be achieved, a myocarditis prediction model with high accuracy is established, and a precious time window is provided for clinicians to formulate a more accurate treatment scheme. The method comprises the following steps: (1) collecting CMR images and clinical data; (2) incorporation and exclusion; (3) pretreatment; (4) sketching an ROI (Region of Interest); (5) dividing a training set and a test set; (6) obtaining a radiomics feature data set; (7) screening; (8) machine learning; (9) utilizing chi-square test and Wilcoxon rank sum test to obtain clinical characteristics with significant difference; (10) splicing the clinical features with significant differences in the step (9) by using the probability value of the prediction model in the step (8), and then modeling a final model by using a logistic regression model; and (11) carrying out model verification on the test set by utilizing the final model.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to an image processing method based on multimodal CMR imaging, an image processing device based on multimodal CMR imaging and applications. Background Art

[0002] Myocarditis can be caused by a variety of microbial infections and is the most common myocardial inflammatory disease. Its incidence is increasing year by year. Some cases progress to dilated cardiomyopathy and then heart failure. It is the most common cause of malignant cardiovascular events in young people. The clinical manifestations of myocarditis vary, from asymptomatic to shortness of breath, severe heart failure to life-threatening ventricular arrhythmias. However, due to the lack of evidence-based clinical medical evidence, it is misdiagnosed or missed, which ultimately delays the best treatment time and leads to disease aggravation or even death. Therefore, it is crucial to establish an efficient method for diagnosing myocarditis.

[0003] Although endomyocardial biopsy remains the gold standard for diagnosing myocarditis and its etiology, it is invasive, costly, and time-consuming. In the field of medical diagnosis, cardiac magnetic resonance imaging is recognized as a non-invasive reference standard for the evaluation of patients with suspected myocarditis. Radiomics, as a tool for feature extraction, analysis, and mining, has been widely used in cancer diagnosis, prognosis assessment, and treatment selection. These technologies (such as support vector machines, decision trees, artificial neural networks, and Bayesian networks) are widely used in cancer research by building predictive models to support effective decision making.

[0004] The latest scientific research progress reveals that imaging omics technology combined with machine learning algorithms has demonstrated significant ability to distinguish myocardial infarction from normal heart status in the application of non-contrast cardiac magnetic resonance imaging, and further successfully distinguished the pathological manifestations of myocardial infarction and myocarditis in the late gadolinium enhancement sequence of magnetic resonance imaging. Despite these advances, the effectiveness of multimodal cardiovascular magnetic resonance imaging (CMR) imaging omics combined with clinical characteristics in the diagnosis of myocarditis still needs to be further explored and verified. Summary of the invention

[0005] In order to overcome the defects of the prior art, the technical problem to be solved by the present invention is to provide an image processing method based on multimodal CMR imaging genomics, which can achieve accurate distinction of cardiovascular magnetic resonance images, establish a highly accurate myocarditis prediction model, and provide clinical physicians with a valuable time window to formulate more accurate treatment plans.

[0006] The technical solution of the present invention is: an image processing method based on multimodal CMR imaging, which comprises the following steps:

[0007] (1) CMR imaging and clinical data collection: Cardiac magnetic resonance imaging scans were performed with the subject in the end-expiratory breath-hold state, combined with the supine vector electrocardiogram gating technology; clinical data of the subject were collected, including identity information, blood routine, comprehensive biochemical tests, coagulation tests, thyroid function tests, and troponin;

[0008] (2) Including and excluding the CMR images and clinical data collected in step (1) according to the specified criteria to obtain the included and excluded CMR images and clinical data sets;

[0009] (3) performing N4 bias field correction, normalization, and image registration on the CMR images and clinical data sets sorted in step (2) to obtain preprocessed CMR images and clinical data sets;

[0010] (4) The cardiac MRI physician delineates the region of interest (ROI) on the preprocessed CMR image dataset to obtain the CMR images and clinical dataset with annotated information;

[0011] (5) stratified sampling of the CMR images and clinical data sets with annotated information in step (4) and divided them into training sets and test sets in a ratio of 8:2;

[0012] (6) For the subjects in the training set and the test set, the PyRadiomics software package was used to extract shape, basic statistics, and complex texture features from the ROI to obtain the radiomics feature dataset;

[0013] (7) For the imaging genomics feature dataset of the training set, feature selection is performed in turn through Pearson correlation coefficient, least absolute shrinkage and selection operator LASSO regression, and principal component analysis PCA dimensionality reduction to obtain the screened imaging genomics feature dataset;

[0014] (8) Using machine learning algorithms such as Gaussian naive Bayes, logistic regression, support vector machine, K-nearest neighbor, gradient boosting, and random forest, the prediction models were established respectively, and the prediction model with the best performance on the test set was retained;

[0015] (9) Using chi-square test and Wilcoxon rank sum test, we obtained clinical data with significant differences.

[0016] feature;

[0017] (10) After using the probability value of the prediction model in step (8) to splice the clinical characteristics with significant differences in step (9), a logistic regression model is used to build the final model;

[0018] (11) On the test set, the final model of step (10) is used to perform model verification. An image processing device based on multimodal CMR imaging is also provided, which includes:

[0019] The data acquisition module is configured to perform cardiac magnetic resonance imaging scans in the subject's end-expiratory breath-hold state in combination with the supine vector electrocardiogram gating technology; collect the subject's clinical data, including: identity information, blood routine, comprehensive biochemical tests, coagulation tests, thyroid function tests, and troponin;

[0020] An inclusion and exclusion module is configured to include and exclude the collected CMR images and clinical data according to specified criteria to obtain the included and excluded CMR images and clinical data sets;

[0021] The preprocessing module is configured to perform N4 bias field correction, normalization and image registration on the collected CMR images and clinical data sets to obtain the preprocessed CMR

[0022] Imaging and clinical datasets;

[0023] The delineation module is configured to allow cardiac MRI doctors to delineate the region of interest (ROI) on the preprocessed CMR image dataset to obtain CMR images and clinical datasets with annotated information;

[0024] A partitioning module is configured to perform stratified sampling on the CMR images and clinical datasets with annotated information and partition them into training and test sets in a ratio of 8:2;

[0025] The feature extraction module is configured to extract shape, basic statistics and complex texture features from ROIs using the PyRadiomics software package for the subjects in the training set and the test set, respectively, to obtain a radiomics feature data set;

[0026] The screening module is configured to select features of the imaging omics feature dataset of the training set by Pearson correlation coefficient, least absolute shrinkage and selection operator LASSO regression, and principal component analysis PCA dimensionality reduction, to obtain the screened imaging omics feature dataset;

[0027] A machine learning module, which is configured to use Gaussian Naive Bayes, Logistic Regression, Support Vector Machine, K-Nearest Neighbor, Gradient Boosting, and Random Forest machine learning algorithms to build prediction models respectively and retain the best prediction model on the test set;

[0028] A clinical characteristics statistical analysis module, which is configured to use chi-square test and Wilcoxon rank sum test to obtain clinical characteristics with significant differences;

[0029] The final model building module is configured to use the probability value of the prediction model in the machine learning module to splice the clinical characteristics with significant differences in the clinical characteristics statistical analysis module, and then use the logistic regression model to build the final model;

[0030] The validation module is configured to perform model validation using the final model on the test set.

[0031] This multimodal CMR imaging-based image processing method is also provided for the prediction of myocarditis.

[0032] The present invention combines the imaging genomics analysis of multimodal CMR and the machine learning model to obtain a basic model for accurately distinguishing patients with myocarditis from patients without myocarditis. Then, the probability values ​​predicted by the basic model and the clinical characteristics with significant differences are spliced ​​to obtain the final prediction model. This can achieve accurate distinction of cardiovascular magnetic resonance images, establish a highly accurate myocarditis prediction model, and provide clinical physicians with a valuable time window to formulate more accurate treatment plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flow chart of an image processing method based on multimodal CMR imaging omics according to the present invention is shown.

[0034] Figure 2 is a schematic diagram of the steps of the method according to the present invention.

[0035] Figure 3 It is a bar chart of the performance of different models on the test set.

[0036] Figure 4 The ROC curves of different models in the task of predicting myocarditis and non-myocarditis for the training set (left) and the test set (right). DETAILED DESCRIPTION

[0037] like Figure 1 , 2 As shown, an image processing method based on multimodal CMR imaging omics comprises the following steps:

[0038] (1) CMR imaging and clinical data collection: Cardiac magnetic resonance imaging scans were performed with the subject in the end-expiratory breath-hold state, combined with the supine vector electrocardiogram gating technology; clinical data of the subject were collected, including identity information, blood routine, comprehensive biochemical tests, coagulation tests, thyroid function tests, and troponin;

[0039] (2) Including and excluding the CMR images and clinical data collected in step (1) according to the specified criteria to obtain the included and excluded CMR images and clinical data sets;

[0040] (3) performing N4 bias field correction, normalization, and image registration on the CMR images and clinical data sets sorted in step (2) to obtain preprocessed CMR images and clinical data sets;

[0041] (4) The cardiac MRI physician delineates the region of interest (ROI) on the preprocessed CMR image dataset to obtain the CMR images and clinical dataset with annotated information;

[0042] (5) stratified sampling of the CMR images and clinical data sets with annotated information in step (4) and divided them into training sets and test sets in a ratio of 8:2;

[0043] (6) For the subjects in the training set and the test set, the PyRadiomics software package was used to extract shape, basic statistics, and complex texture features from the ROI to obtain the radiomics feature dataset;

[0044] (7) For the imaging genomics feature dataset of the training set, feature selection is performed in turn through Pearson correlation coefficient, least absolute shrinkage and selection operator LASSO regression, and principal component analysis PCA dimensionality reduction to obtain the screened imaging genomics feature dataset;

[0045] (8) Using machine learning algorithms such as Gaussian naive Bayes, logistic regression, support vector machine, K-nearest neighbor, gradient boosting, and random forest, the prediction models were established respectively, and the prediction model with the best performance on the test set was retained;

[0046] (9) Using chi-square test and Wilcoxon rank sum test, clinical characteristics with significant differences were obtained;

[0047] (10) After using the probability value of the prediction model in step (8) to splice the clinical characteristics with significant differences in step (9), a logistic regression model is used to build the final model;

[0048] (11) On the test set, use the final model of step (10) to perform model verification.

[0049] The present invention combines the imaging genomics analysis of multimodal CMR and the machine learning model to obtain a basic model for accurately distinguishing patients with myocarditis from patients without myocarditis. Then, the probability values ​​predicted by the basic model and the clinical characteristics with significant differences are spliced ​​to obtain the final prediction model. This can achieve accurate distinction of cardiovascular magnetic resonance images, establish a highly accurate myocarditis prediction model, and provide clinical physicians with a valuable time window to formulate more accurate treatment plans.

[0050] Preferably, in step (1), during CMR image acquisition, basic cardiac data are first acquired through a multi-position spin echo sequence; then, a balanced turbine field echo cine sequence BTFE is used to scan the long axis of two chambers, three chambers, four chambers and continuous short axis images of the left ventricle; then, an inversion recovery gradient echo PSIR sequence is used to acquire long axis and short axis images of the left ventricle; finally, two radiologists independently measure the left ventricular ejection fraction, stroke volume, cardiac output, end-diastolic volume and end-systolic volume from the CMR images.

[0051] Preferably, in step (2), the inclusion criteria for myocarditis patients include: patients must show clinical symptoms and signs related to myocarditis, and the course of the disease meets the definition of the disease; serum cardiac troponin is elevated; CMR examination shows at least one T2-related sequence and at least one T1-related sequence are simultaneously positive; Exclusion criteria: combined congenital heart disease, rheumatic heart disease, hyperthyroidism heart disease; CMR image quality does not meet diagnostic requirements, affecting the assessment of myocarditis; there is a lack of key clinical data;

[0052] The inclusion criteria for patients without myocarditis included: no history of myocarditis or other known serious heart diseases; complete imaging data available for analysis. Exclusion criteria included: individuals with a family history of heart disease, hypertension, and hyperlipidemia; CMR image quality that did not meet diagnostic requirements; and missing key clinical data.

[0053] Preferably, in step (3), the CMR images are selected from three sequences of cardiac magnetic resonance, namely, the movie sequence BTFE, T2-weighted imaging T2WI and myocardial gadolinium delayed enhancement imaging LGE.

[0054] Preferably, in step (3), the CMR images and clinical data sets are subjected to N4 bias field correction, normalization and standardization processing on all images using the Python language, and then all images of each enrolled myocarditis patient are co-registered with the movie images to obtain the preprocessed CMR images and clinical data sets.

[0055] Preferably, in step (6), for the training set and the test set, the PyRadiomics software package is used to extract 14 shape features, 18 first-order statistics, 24 grayscale co-occurrence matrix features, 16 grayscale size region matrix features, 16 grayscale run length matrix features, 5 adjacent grayscale difference matrix features and 14 grayscale dependency matrix features from the ROI, respectively, to obtain an imaging genomics feature data set.

[0056] Preferably, in step (7), for the imaging genomics feature dataset of the training set, the Pearson correlation coefficient is used to evaluate the features pairwise, and a strict threshold of 0.75 is set to identify and remove highly correlated redundant features; then, a LASSO regression model based on GridSearchCV 5-fold cross-validation is used for feature screening; then, PCA dimensionality reduction is performed to map the features screened by LASSO to a low-dimensional space that retains the maximum data variability, and the top 10 principal components are determined to obtain the screened imaging genomics feature dataset.

[0057] Preferably, in step (11), the performance of each model is comprehensively measured based on evaluation indicators such as area under the curve (AUC), accuracy, recall, precision and F1 score.

[0058] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes the steps of the above-mentioned embodiment method, and the storage medium can be: ROM / RAM, disk, CD, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes an image processing device based on multimodal CMR imaging omics, which is usually represented in the form of functional modules corresponding to the steps of the method. The device includes:

[0059] The data acquisition module is configured to perform cardiac magnetic resonance imaging scans in the subject's end-expiratory breath-hold state in combination with the supine vector electrocardiogram gating technology; collect the subject's clinical data, including: identity information, blood routine, comprehensive biochemical tests, coagulation tests, thyroid function tests, and troponin;

[0060] An inclusion and exclusion module is configured to include and exclude the collected CMR images and clinical data according to specified criteria to obtain the included and excluded CMR images and clinical data sets;

[0061] A preprocessing module configured to sequentially perform N4 bias field correction, normalization, and image registration on the sorted CMR images and clinical data sets to obtain preprocessed CMR images and clinical data sets;

[0062] The delineation module is configured to allow cardiac MRI doctors to delineate the region of interest (ROI) on the preprocessed CMR image dataset to obtain CMR images and clinical datasets with annotated information;

[0063] A partitioning module is configured to perform stratified sampling on the CMR images and clinical datasets with annotated information and partition them into training and test sets in a ratio of 8:2;

[0064] The feature extraction module is configured to extract shape, basic statistics and complex texture features from ROIs using the PyRadiomics software package for the subjects in the training set and the test set, respectively, to obtain a radiomics feature data set;

[0065] The screening module is configured to select features of the imaging omics feature dataset of the training set by Pearson correlation coefficient, least absolute shrinkage and selection operator LASSO regression, and principal component analysis PCA dimensionality reduction, to obtain the screened imaging omics feature dataset;

[0066] A machine learning module, which is configured to use Gaussian Naive Bayes, Logistic Regression, Support Vector Machine, K-Nearest Neighbor, Gradient Boosting, and Random Forest machine learning algorithms to build prediction models respectively and retain the best prediction model on the test set;

[0067] A clinical characteristics statistical analysis module, which is configured to use chi-square test and Wilcoxon rank sum test to obtain clinical characteristics with significant differences;

[0068] The final model building module is configured to use the probability value of the prediction model in the machine learning module to splice the clinical characteristics with significant differences in the clinical characteristics statistical analysis module, and then use the logistic regression model to build the final model;

[0069] The validation module is configured to perform model validation using the final model on the test set.

[0070] This multimodal CMR imaging-based image processing method is also provided for the prediction of myocarditis.

[0071] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings. The method of the present invention comprises:

[0072] Step 1: CMR imaging and clinical data collection. A total of 107 myocarditis patients and 104 non-myocarditis volunteers were obtained from our hospital. The following clinical variables of myocarditis patients and non-myocarditis volunteers were collected:

[0073] (a) CMR imaging data acquisition: During each patient's hospitalization, CMR imaging scans were performed using a 3.0-T MRI system in accordance with the standardized guidelines of the Society of Cardiovascular Magnetic Resonance, with the patient in the end-expiratory breath-hold state, combined with supine vector electrocardiogram gating technology. First, basic cardiac data were acquired using a conventional multi-position spin echo (SE) sequence; then, a balanced turbine field echo cine sequence (BTFE) was used to obtain long-axis images of the two-chamber, three-chamber, and four-chamber heart and continuous short-axis images of the left ventricle (parameters: repetition time (TR): 2.9 ms; echo time (TE): 1.47 ms; flip angle (FA): 45; slice thickness: 8 mm). Finally, to evaluate myocardial fibrosis and scar, 10 minutes after the injection of 0.3mmol / kg gadolinium amide contrast agent, left ventricular long-axis and short-axis images were acquired using an inversion recovery gradient echo sequence (parameters: flip angle 25°, TR / TE 6 / 3ms, TI 260-350ms, voxel 1.6×1.9×8mm3, acquisition time 100-125ms). Two experienced radiologists independently interpreted the CMR images and measured key cardiac function indicators such as left ventricular ejection fraction, stroke volume, cardiac output, end-diastolic volume, and end-systolic volume.

[0074] (b) Clinical data collection: patients’ general information (sex, age, height, weight, systolic blood pressure, diastolic blood pressure, history of diabetes, history of hypertension, smoking history, drinking history) and laboratory test information (blood routine, comprehensive biochemical test, coagulation test, thyroid function test, troponin);

[0075] Step 2, inclusion and exclusion. Based on the latest guidelines for the diagnosis and treatment of myocarditis in JCS in 2023, the inclusion criteria for myocarditis include: (a) patients must show clinical symptoms and signs related to myocarditis, such as chest pain, palpitations, dyspnea, etc., and the course of the disease meets the definition of myocarditis; (b) serum cardiac troponin (cTn) is elevated, especially persistent elevation of high-sensitivity cTn, indicating myocardial damage; (c) CMR examination must show at least one T2-related sequence and at least one T1-related sequence simultaneously positive to prove myocardial edema or other signs of myocardial damage; (d) no age limit, no gender limit. Exclusion criteria: (a) Other heart diseases: patients with known heart diseases such as congenital heart disease, rheumatic heart disease, hyperthyroid heart disease, etc.; (b) CMR image quality does not meet diagnostic requirements, affecting the evaluation of myocarditis; (c) There is a lack of key clinical data. For the inclusion of non-myocarditis volunteers, the following criteria were set: (a) no history of myocarditis and other known serious heart diseases; (b) in terms of age and gender distribution, try to match the myocarditis patient group to reduce potential bias; (c) have complete imaging data that can be used for analysis. Exclusion criteria: (a) individuals with clear risk factors for heart disease such as family history of heart disease, hypertension, and hyperlipidemia will be carefully considered for inclusion to reduce potential interference with the study results; (b) CMR image quality does not meet diagnostic requirements; (d) there is a lack of key clinical data. Finally, this study included 54 myocarditis patients and 65 non-myocarditis volunteers who were hospitalized in our hospital from January 1, 2019 to June 30, 2024.

[0076] Step 3, preprocessing. According to the latest Lake Louise criteria for myocarditis, this study selected three sequences of cardiac magnetic resonance imaging, namely movie sequence (BTFE), T2-weighted imaging (T2WI), and myocardial gadolinium delayed enhancement imaging (LGE). All images were corrected for N4 bias field, normalized, and standardized using the python language (version 3.11), and then all images of each enrolled patient were co-registered with the movie image.

[0077] Step 4: Delineate the region of interest (ROI). Finally, two experienced cardiac MRI experts used 3D Slicer software (version 5.6.1) to manually segment the left ventricular myocardium of each enrolled subject on the short-axis movie image and defined it as the ROI, obtaining CMR images and clinical data sets with annotation information.

[0078] Step 5: Dataset division: The CMR images and clinical datasets with annotated information in step 4 are divided into training and test sets using stratified sampling in a ratio of 8:2.

[0079] Step 6, extraction of radiomics features. For patients / volunteers in the training set and test set, 14 shape features, 18 first-order statistics, 24 grayscale co-occurrence matrix features, 16 grayscale size area matrix features, 16 grayscale run length matrix features, 5 adjacent grayscale difference matrix features, and 14 grayscale dependency matrix features were extracted from ROI using the PyRadiomics package to obtain the radiomics feature dataset.

[0080] Step 7, feature selection. For the imaging omics feature dataset of the training set, the Pearson correlation coefficient was used to evaluate the features pairwise, and a strict threshold of 0.75 was set to identify and remove highly correlated redundant features. The LASSO regression model based on GridSearchCV 5-fold cross-validation was then used for feature screening. Finally, the PCA dimensionality reduction technique was used to map the features screened by LASSO to a low-dimensional space that retains the maximum data variability, and the top 10 principal components were determined to obtain the screened imaging omics feature dataset.

[0081] Step 8: Model building. Use Gaussian naive Bayes, logistic regression, support vector machine, K-nearest neighbor, gradient boosting and random forest to build prediction models. Figure 3 As shown, the best performing model on the test set is retained.

[0082] Step 9: Statistical analysis of clinical characteristics: Python 3.7.13 was used to perform the analysis. The chi-square test and WilcoxonMann-Whitney U test were used to obtain 35 clinical characteristics with significant differences, as shown in Table 1.

[0083] Step 10: Final model building: After the probability value of the prediction model in step 8 is combined with the clinical features with significant differences in step 9, the final model building is performed using a logistic regression model.

[0084] Step 11: Model validation. On the test set, the final model of step 10 is used for model validation. Specifically, the area under the curve (AUC) is used to calculate the model. Figure 4 The performance of each model is comprehensively measured by using multi-dimensional evaluation indicators such as accuracy, recall, precision, and F1 score. The results are shown in Table 2. The accuracy of this model on the test set is 0.917, and the AUC is 0.944, achieving the best results.

[0085] Table 1 35 clinical characteristics with significant differences obtained from statistical analysis.

[0086]

[0087]

[0088]

[0089] Note: a: Mann-Whitney-U test; b: Chi-square test.

[0090] Table 2 Performance of seven machine learning models in myocarditis and non-myocarditis tasks. The last random forest + LR is a model with significantly different clinical features added.

[0091]

[0092] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the protection scope of the technical solution of the present invention.

Claims

1. An image processing method based on multimodal CMR radiomics, characterized by: It includes the following steps: (1) CMR imaging and clinical data acquisition: Cardiac magnetic resonance imaging scans were performed with the subject in the end-expiratory breath-hold state using the supine vector electrocardiogram gating technique; Collect clinical data of the subjects, including: identity information, blood routine, comprehensive biochemical test, coagulation test, thyroid function test, troponin; (2) Including and excluding the CMR images and clinical data collected in step (1) according to the specified criteria to obtain the included and excluded CMR images and clinical data sets; (3) performing N4 bias field correction, normalization, and image registration on the CMR images and clinical data sets sorted in step (2) to obtain preprocessed CMR images and clinical data sets; (4) The cardiac MRI physician delineates the region of interest (ROI) on the preprocessed CMR image dataset to obtain the CMR images and clinical dataset with annotated information; (5) stratified sampling of the CMR images and clinical data sets with annotated information in step (4) and divided them into training sets and test sets in a ratio of 8:2; (6) For the subjects in the training set and the test set, the PyRadiomics software package was used to extract shape, basic statistics, and complex texture features from the ROI to obtain the radiomics feature dataset; (7) For the imaging genomics feature dataset of the training set, feature selection is performed in turn through Pearson correlation coefficient, least absolute shrinkage and selection operator LASSO regression, and principal component analysis PCA dimensionality reduction to obtain the screened imaging genomics feature dataset; (8) Using machine learning algorithms such as Gaussian naive Bayes, logistic regression, support vector machine, K-nearest neighbor, gradient boosting, and random forest, the prediction models were established respectively, and the prediction model with the best performance on the test set was retained; (9) Using chi-square test and Wilcoxon rank sum test, clinical characteristics with significant differences were obtained; (10) After using the probability value of the prediction model in step (8) to splice the clinical characteristics with significant differences in step (9), a logistic regression model is used to build the final model; (11) On the test set, use the final model of step (10) to perform model verification.

2. The image processing method based on multimodal CMR radiomics according to claim 1, characterized in that: In the step (1), when acquiring CMR images, basic cardiac data are first acquired through a multi-position spin echo sequence; then, a balanced turbine field echo cine sequence BTFE is used to scan the long axis of the two chambers, three chambers, four chambers and continuous short axis images of the left ventricle; then, an inversion recovery gradient echo PSIR sequence is used to acquire long axis and short axis images of the left ventricle; finally, two radiologists independently measure the left ventricular ejection fraction, stroke volume, cardiac output, end-diastolic volume and end-systolic volume from the CMR images.

3. The image processing method based on multimodal CMR radiomics according to claim 2, characterized in that: In step (2), the inclusion criteria for myocarditis patients include: patients must show clinical symptoms and signs related to myocarditis, and the course of the disease meets the definition of the disease; serum cardiac troponin is elevated; CMR examination shows at least one T2-related sequence and at least one T1-related sequence are simultaneously positive; Exclusion criteria: congenital heart disease, rheumatic heart disease, hyperthyroidism heart disease; CMR image quality does not meet diagnostic requirements, affecting the assessment of myocarditis; there is a lack of key clinical data; Inclusion criteria for non-myocarditis patients included: no history of myocarditis or other known serious heart diseases; complete imaging data available for analysis; exclusion criteria included: individuals with a family history of heart disease, hypertension, and hyperlipidemia; CMR image quality that did not meet diagnostic requirements; and missing key clinical data.

4. The image processing method based on multimodal CMR radiomics according to claim 3, characterized in that: In the step (3), the CMR images are selected from three sequences of cardiac magnetic resonance, namely, the movie sequence BTFE, T2 weighted imaging T2WI and myocardial gadolinium delayed enhancement imaging LGE.

5. The image processing method based on multimodal CMR radiomics according to claim 4, characterized in that: In the step (3), the CMR images and clinical data sets are subjected to N4 bias field correction, normalization and standardization processing using Python language, and then all images of each enrolled myocarditis patient are co-registered with the movie images to obtain the preprocessed CMR images and clinical data sets.

6. The image processing method based on multimodal CMR radiomics according to claim 5, characterized in that: In the step (6), for the training set and the test set, the PyRadiomics software package is used to extract 14 shape features, 18 first-order statistics, 24 grayscale co-occurrence matrix features, 16 grayscale size region matrix features, 16 grayscale run length matrix features, 5 adjacent grayscale difference matrix features and 14 grayscale dependency matrix features from the ROI, respectively, to obtain an imaging genomics feature data set.

7. The image processing method based on multimodal CMR radiomics according to claim 6, characterized in that: In step (7), for the radiomics feature dataset of the training set, the Pearson correlation coefficient is used to evaluate the features pairwise, and a strict threshold of 0.75 is set to identify and remove highly correlated redundant features; Subsequently, based on The LASSO regression model with GridSearchCV 5-fold cross-validation was used for feature screening. Then, PCA dimensionality reduction was used to map the features screened by LASSO into a low-dimensional space that retained the maximum data variability, and the top 10 principal components were determined to obtain the screened imaging genomics feature dataset.

8. The image processing method based on multimodal CMR radiomics according to claim 7, characterized in that: In the step (11), the performance of each model is comprehensively measured based on the evaluation indicators of area under the curve AUC, accuracy, recall, precision and F1 score.

9. An image processing device based on multimodal CMR radiomics, characterized in that: It includes: A data acquisition module configured to perform cardiac magnetic resonance imaging scanning in a subject's end-expiratory breath-hold state in combination with a supine vector electrocardiogram gating technique; Collect clinical data of the subjects, including: identity information, blood routine, comprehensive biochemical test, coagulation test, thyroid function test, troponin; An inclusion and exclusion module is configured to include and exclude the collected CMR images and clinical data according to specified criteria to obtain the included and excluded CMR images and clinical data sets; A preprocessing module configured to sequentially perform N4 bias field correction, normalization, and image registration on the sorted CMR images and clinical data sets to obtain preprocessed CMR images and clinical data sets; The delineation module is configured to allow cardiac MRI doctors to delineate the region of interest (ROI) on the preprocessed CMR image dataset to obtain CMR images and clinical datasets with annotated information; A partitioning module is configured to perform stratified sampling on the CMR images and clinical datasets with annotated information and partition them into training and test sets in a ratio of 8:2; The feature extraction module is configured to extract shape, basic statistics and complex texture features from ROIs using the PyRadiomics software package for the subjects in the training set and the test set, respectively, to obtain a radiomics feature data set; The screening module is configured to select features of the imaging omics feature dataset of the training set by Pearson correlation coefficient, least absolute shrinkage and selection operator LASSO regression, and principal component analysis PCA dimensionality reduction, to obtain the screened imaging omics feature dataset; A machine learning module, which is configured to use Gaussian Naive Bayes, Logistic Regression, Support Vector Machine, K-Nearest Neighbor, Gradient Boosting, and Random Forest machine learning algorithms to build prediction models respectively and retain the best prediction model on the test set; A clinical characteristics statistical analysis module, which is configured to use chi-square test and Wilcoxon rank sum test to obtain clinical characteristics with significant differences; The final model building module is configured to use the probability value of the prediction model in the machine learning module to splice the clinical characteristics with significant differences in the clinical characteristics statistical analysis module, and then use the logistic regression model to build the final model; The validation module is configured to perform model validation using the final model on the test set.

10. The application of the image processing method based on multimodal CMR radiomics according to claim 1, characterized in that: It is used for the prediction of myocarditis.

Images