Cardiovascular disease detection method and system fusing medical image and molecular omics characteristics

Abstract

The invention discloses a cardiovascular disease detection method and system fusing medical imaging and molecular omics characteristics. The method comprises the following steps of: obtaining data required by a model, wherein the data required by the model is listed in details subsequently; training a model by using the data, and testing and verifying; the output result of the cardiovascular disease risk prediction model is the cardiovascular disease risk and possible risk characteristics of the data source sample. The method is based on a large-scale crowd queue, and has a large amount of related clinical information and physiological indexes of 500,000 samples; according to the method, a classical ResNet network model and a basic neural network model are used for constructing a multi-modal model, and connection between different modal data is established. The risk of cardiovascular diseases of individuals is detected from the aspects of images, genes and metabolism.

Description

Technical Field

[0001] The present invention relates to a cardiovascular disease detection method and system integrating medical imaging and molecular omics features, and also to a corresponding model for cardiovascular disease risk prediction and risk feature description based on imaging omics, belonging to the field of artificial intelligence-assisted diagnosis technology. Background Art

[0002] Cardiovascular disease is a general term for a group of diseases that affect the heart and vascular system, mainly including coronary heart disease, stroke, myocardial infarction, heart failure, hypertension, etc. These diseases are usually caused by factors such as arteriosclerosis (the accumulation of fat, cholesterol and other substances on the inner wall of the artery to form plaques, causing blood vessels to narrow or block), thrombosis and reduced vascular elasticity, which in turn affect the circulation of blood in the heart and various parts of the body. Cardiovascular disease is one of the leading causes of death worldwide. Its occurrence and development are related to a variety of risk factors, such as smoking, unhealthy diet, lack of exercise, obesity, diabetes, high blood pressure, high cholesterol, genetic factors and age.

[0003] Cardiovascular disease not only seriously threatens the life and health of patients, but also brings a heavy burden to families and society. It will lead to a decline in the quality of life of patients, increase family economic pressure, and occupy a large amount of medical resources. As one of the leading causes of death in the world, cardiovascular disease poses a huge challenge to the public health system and socioeconomic development. Early prediction and detection of cardiovascular disease can help identify high-risk groups early, so as to intervene in time and avoid the occurrence of serious complications as much as possible. Through prediction, patients can take preventive measures, adjust their lifestyles or receive early treatment before the disease progresses, which not only improves the efficiency of health management, but also reduces the consumption of medical resources. Although there are currently a variety of prediction methods for cardiovascular disease available, there are problems such as insufficient accuracy, high cost and narrow applicability. Summary of the invention

[0004] The primary technical problem to be solved by the present invention is to provide a method for predicting cardiovascular disease risk and describing risk characteristics based on medical imaging and molecular omics.

[0005] Another technical problem to be solved by the present invention is to provide a system for predicting cardiovascular disease risk and describing risk characteristics based on medical imaging and molecular omics.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention first discloses a cardiovascular disease detection method integrating medical imaging and molecular omics features, which specifically includes:

[0008] Step S1, obtaining data to be collected, including aortic dilatation image data, cardiac long-axis image data, cardiac short-axis image data, metabolomics data, and genomics data of the subject;

[0009] Step S2, inputting the processed data into a cardiovascular disease risk prediction model for training, testing and verification;

[0010] Step S3, the output result of the cardiovascular disease risk prediction model is the disease probability of the subject to be investigated;

[0011] Step S4: Calculate the SHAP value, analyze the risk characteristic status of the person to be investigated, and visualize it.

[0012] Preferably, when obtaining the cardiac aorta dilatation image data, the following processing is performed:

[0013] Step 111, in the aortic dilatation Magnetic Resonance Imaging (MRI) images of 78684 people, convert the above image sequence into NIFTI format according to the image details; and unify the image selection into the data of instance2.

[0014] Preferably, when obtaining the cardiac long-axis image data, the following processing is performed:

[0015] Step 121, among the cardiac long-axis MRI images of 83698 persons, convert the above image sequence into NIFTI format according to the image details; in the selection of the long-axis NIFTI image, select the 2ch image.

[0016] Preferably, when obtaining the cardiac short-axis image data, the following processing is performed:

[0017] Step 131: among the cardiac short-axis MRI images of 83,469 persons, the above image sequence is converted into the NIFTI format according to the image details; and the image selection is unified as the data of instance 2.

[0018] Preferably, when obtaining the metabolomics data, the following processing is performed:

[0019] Step 141: First, the metabolomics data of 502,225 people were preprocessed, including quality control and normalization. Then, the Cox proportional hazard model was used to study the longitudinal association between metabolites and cardiovascular disease. The importance of these metabolites was quantified using machine learning methods, and then the metabolic risk score (MetRs) of each sample was calculated.

[0020] Preferably, when the genomics data is obtained, the following processing is performed:

[0021] Step 121, first collect summary statistics of relevant cardiovascular diseases, then use PLINK software to perform standardized quality control on the genomic data of 488,377 people to eliminate unqualified samples and SNPs, and finally use PRSice-2 software to calculate the polygenic risk score of each sample.

[0022] Preferably, the original characteristic imaging data collected by S1 are sequentially subjected to data quality control and format conversion; the original characteristic metabolomics data collected by S1 are sequentially subjected to data deletion processing and quality control; the genomic data collected by S1 are subjected to format conversion and phenotype filling; quality control refers to excluding individuals with poor sample quality; format conversion refers to processing into a required format according to the requirements of specific software, such as DICOM to NIFTI, bed / bim / fam to map / ped; deletion processing refers to filtering out data with a data deletion rate greater than 1%; phenotype filling refers to filling in phenotype values ​​according to whether the individual is sick or not, where not sick = 1 and sick = 2;

[0023] The cardiovascular disease risk prediction model is a multimodal model consisting of an imaging model and a genetic model. It is mainly obtained through the following steps:

[0024] Step S21, obtaining multiple groups of survey data of normal and cardiovascular disease patients, each group of survey data is aortic dilatation imaging data, cardiac long-axis imaging data, cardiac short-axis imaging data, metabolome data, and genomic data of a certain respondent after a standardized processing process; and using a multimodal model to screen out relevant cardiovascular features from these data for modeling, and feature selection is carried out in multiple steps, as follows:

[0025] First, in the imaging model part, the original feature image data (NIFTI format) is uniformly converted into the input format of the ResNet50 model; the professional medical image processing software SimpleITK is used to extract 3D slices and convert them into 2D images, and then the images are normalized and resized to adapt to the input format of the imaging model;

[0026] Then, a pre-trained ResNet50 model is used to extract features from the two-dimensional image of the heart, and feature vectors are extracted from the three two-dimensional views of the heart. The feature vectors of the three views are concatenated or superimposed to form a joint feature representation of the entire heart. The principal component analysis method (PCA) is used to reduce the dimension of the extracted features, reduce unimportant features, and output the feature vectors to the embedding space.

[0027] Step S22, obtaining multiple groups of survey data on normal subjects and subjects with cardiovascular diseases, each group of survey data being genomic data of a certain subject after a standardized processing flow; selecting features from these genomic data to establish a genetic model, mainly using a neural network model to process the genomic data through a fully connected layer, normalization, and activation function, and then outputting the feature results to an embedding space;

[0028] Step S23, obtaining multiple groups of survey data on normal and cardiovascular disease patients, each group of survey data is the metabolomics data of a certain respondent after a standardized processing flow; screening features from these metabolomics data to establish a genetic model, mainly using a neural network model to process the metabolomics data through a fully connected layer, normalization and activation function, and then outputting the feature results to the embedding space.

[0029] Step S24: Use the above image model and genetic model as input models of the multimodal model, establish a neural network model for comparative learning, process the joint embedding of image data, gene data and metabolic data, and the final output is a binary classification result.

[0030] Step S25, training a pre-designed cardiovascular disease risk prediction model with the above data of multiple groups of normal people and people with cardiovascular diseases, and further screening variables by multiple cross-validation + deep learning to obtain the optimal cardiovascular disease risk prediction model;

[0031] The cardiovascular disease risk prediction model is superior in that it uses a multimodal neural network model that combines an imaging model with a genetic model. For the imaging model part, the classic ResNet50 model structure is used. The entire network structure includes an input layer of ResNet50, 16 residual blocks, and a pooling layer.

[0032] The genetic model: maps the input genome / metabolome data through two fully connected layers and a ReLu activation function to map the genetic features / metabolic features and output a feature vector;

[0033] Preferably, after making a prediction, the cardiovascular disease risk prediction model uses the SHAP method for interpretation, during which a SHAP model interpreter is constructed, and the average value of the training data is used as the background value; the SHAP value of each sample is calculated, and the influence of the features of the images of different views on the model prediction results is analyzed to evaluate their contribution values; finally, a SHAP interpretation diagram is generated to intuitively show how the features of the aorta, long axis, short axis, genome and metabolome affect the model's prediction of heart health risks;

[0034] According to a second aspect of an embodiment of the present invention, a multimodal model for cardiovascular disease risk prediction and risk feature description based on radiomics is provided, including an imaging model, a genetic model, and a result display. The imaging model is combined with the genetic model for comparative learning, and finally a prediction result is obtained, which is used to perform the following operations:

[0035] Obtaining data to be collected, including the subject's aortic dilatation imaging data, cardiac long-axis imaging data, cardiac short-axis imaging data, metabolomics data, and genomics data;

[0036] Inputting the processed data into a cardiovascular disease risk prediction model;

[0037] The output results of the cardiovascular disease risk prediction model are the probability of illness of the subject to be investigated and the ranking of risk characteristics.

[0038] The present invention also discloses a cardiovascular disease detection system that integrates medical imaging and molecular genomics features, wherein the system comprises at least one processor and a memory connected to the processor, wherein: the memory is used to store computer programs or instructions; the processor is used to obtain and execute the computer programs or instructions in the memory, so that the system executes the detection method disclosed in the present invention.

[0039] Beneficial effects of the present invention

[0040] The method for predicting and characterizing cardiovascular disease risk based on radiomics provided by the present invention extracts shape, texture and statistical features from medical images, combined with machine learning models, to accurately assess the patient's disease risk and identify key risk features. This method can not only detect high-risk individuals at an early stage and guide personalized intervention measures, but also provide doctors with intuitive risk feature explanations to assist clinical decision-making. Its beneficial effects include improving diagnostic efficiency, reducing missed diagnoses, and significantly optimizing the allocation of medical resources, promoting the application of precision medicine in the prevention and control of cardiovascular diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of the process of constructing a cardiovascular disease risk prediction model provided by an embodiment of the present invention

[0042] Figure 2 Schematic diagram of the processing flow of data used in the present invention

[0043] Figure 3 The cardiovascular disease risk prediction model constructed in the present invention is

[0044] Figure 4 ROC curve of the genetic model constructed in the present invention

[0045] Figure 5The risk characteristics of the genetic model constructed in the present invention DETAILED DESCRIPTION

[0046] The present invention will be further described below in conjunction with embodiments, but the protection scope of the present invention is not limited thereto:

[0047] For people who are suspected of having cardiovascular disease, in order to help doctors predict the risk of patients (or help patients self-diagnose) for cardiovascular disease, such as Figure 1 As shown, the method for predicting cardiovascular disease risk based on radiomics provided in an embodiment of the present invention comprises the following steps:

[0048] Step S1, obtaining the data to be collected, including the aortic dilatation imaging data, cardiac long-axis imaging data, cardiac short-axis imaging data, metabolomics data, and genomics data of the subject. We collect features on a large scale to increase the accuracy of the model, and the data processing flow is as follows: Figure 2 shown.

[0049] When obtaining the cardiac aorta dilatation image data, the following steps are included:

[0050] Step 111, in the aortic dilatation Magnetic Resonance Imaging (MRI) images of 78684 people, convert the above image sequence into NFITI format according to the image details; and unify the image selection into the data of instance2.

[0051] When obtaining the cardiac long-axis image data, the following processing is performed, including the following processing:

[0052] Step 121, among the cardiac long-axis MRI images of 83698 persons, convert the above image sequence into NFITI format according to the image details; in the selection of the long-axis NFITI image, select the 2ch image.

[0053] When obtaining the cardiac short-axis image data, the following steps are included:

[0054] Step 131: among the cardiac short-axis MRI images of 83,469 persons, the above image sequence is converted into the NFITI format according to the image details; and the image selection is unified as the data of instance 2.

[0055] When obtaining the metabolomics data, the following steps are included:

[0056] Step 141: First, the metabolomics data of 502,225 people were preprocessed, including quality control and normalization. Then, the Cox proportional hazard model was used to study the longitudinal association between metabolites and cardiovascular disease. The importance of these metabolites was quantified using machine learning methods. Then, the metabolic risk score (MetRs) of each sample was calculated. Figure 4 The ROC curve in Figure 2 shows that the accuracy of the genetic model in predicting cardiovascular disease is roughly equivalent to a 70% confidence level. This means that the genetic model can predict cardiovascular disease to a certain extent based on metabolites, and the trend of the ROC curve further shows the sensitivity and specificity of the model at different judgment thresholds.

[0057] When obtaining the genomics data, the following processing is included:

[0058] Step 121, first collect summary statistics of relevant cardiovascular diseases, then use PLINK software to perform standardized quality control on the genomic data of 488,377 people to eliminate unqualified samples and SNPs, and finally use PRSice-2 software to calculate the polygenic risk score of each sample.

[0059] Step S2: input the processed data into a pre-trained cardiovascular disease risk prediction model.

[0060] The pre-trained cardiovascular disease risk prediction model is obtained through the following steps:

[0061] Step S21, obtaining multiple groups of survey data on normal people and people with cardiovascular diseases, each group of survey data is the heart aortic dilation imaging data, heart long-axis imaging data, heart short-axis imaging data, metabolome data, and genome data of a certain respondent after a standardized processing flow; and using the established multimodal model to screen out relevant cardiovascular characteristics from these data for modeling.

[0062] The method for obtaining the data to be investigated in step S1 is adopted to obtain the above-mentioned investigation data of multiple groups of normal patients and patients with cardiovascular diseases, which are used to train the cardiovascular disease risk prediction model.

[0063] Step S22, obtaining multiple groups of the above data of normal subjects and subjects suffering from cardiovascular diseases to train a pre-designed cardiovascular disease risk prediction model, and further screening variables by multiple cross-validation + deep learning to obtain the optimal cardiovascular disease risk prediction model.

[0064] The cardiovascular disease risk prediction model adopts a multimodal pre-trained neural network model that combines an imaging model with a genetic model. For the imaging model part, the ResNet50 classic model with the last fully connected layer removed is used. The entire network structure includes an input layer, a residual layer, and an average pooling layer. For the genetic model, a simple neural network model is used, including two fully connected layers and a ReLu activation function.

[0065] Step S3: The output result of the cardiovascular disease risk prediction model is the disease probability of the person to be investigated.

[0066] Step S4: Calculate the SHAP value, analyze the risk characteristic status of the person to be investigated and visualize it, such as Figure 5 As shown, SHAP software was used to quantify the contribution of each metabolite to the prediction results.

[0067] The collected data obtained in step S1 is input into a pre-trained cardiovascular disease risk prediction model, which will output the probability of the subject suffering from cardiovascular disease in the range of 0-1. This is to assist the medical system in making a diagnosis based on the subject's probability of disease and disease conditions, or the subject uses the model for pre-screening. If the output probability of this model is greater than 50%, it means that the subject is more likely to suffer from cardiovascular disease.

[0068] To implement the method for describing cardiovascular disease and risk characteristics based on radiomics provided by the present invention, Figure 3 As shown, the present invention also provides a model for cardiovascular disease risk prediction and risk characteristic description based on imaging omics.

[0069] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A cardiovascular disease detection method integrating medical imaging and molecular omics features, characterized in that It includes the following steps: Step S1, obtaining and preprocessing the data to be collected, wherein the data to be collected includes original characteristic imaging data, metabolomics data, and genomics data of the subject; Step S2: input the processed data into a pre-trained cardiovascular disease risk prediction model; the cardiovascular disease risk prediction model is composed of an image model and a genetic model, and the cardiovascular disease risk prediction model is obtained through the following steps: Step S21, obtaining multiple groups of normal and cardiovascular disease imaging survey data, each group of survey data is the original feature image data of a certain subject after preprocessing; and using the ResNet model to filter out relevant cardiovascular features from the original feature image data to establish an image model; Step S22, obtaining multiple groups of survey data of normal people and people with cardiovascular diseases, each group of survey data being pre-processed genomic data of a certain person under investigation; screening features from the genomic data to establish a genetic model; Step S23, obtaining multiple groups of survey data of normal subjects and subjects with cardiovascular diseases, each group of survey data being pre-processed metabolomics data of a certain subject; screening features from the metabolomics data to establish a genetic model; Step S24: Use the above image model and genetic model as input models of the multimodal model, establish a neural network model for contrastive learning, process the joint embedding of image data, gene data and metabolic data, and the final output is a binary classification result; Step S25, obtaining a cardiovascular disease risk prediction model designed by training the above data of multiple groups of normal people and people with cardiovascular diseases, and further screening variables by multiple cross-validation + deep learning to obtain the optimal cardiovascular disease risk prediction model as the pre-trained cardiovascular disease risk prediction model described in S2; Step S3: The output result of the cardiovascular disease risk prediction model is the disease probability of the person to be investigated.

2. The method according to claim 1, characterized in that The original characteristic image data includes cardiac aortic dilatation image data, cardiac long-axis image data, and cardiac short-axis image data.

3. The method according to claim 1, characterized in that The pre-processing specifically includes: Perform data quality control and format conversion on the original feature image data in sequence; The metabolomics data were processed for missing information and quality control in turn; Perform format conversion, phenotype filling, and quality control on genomics data.

4. The method according to claim 3, characterized in that: Obtaining cardiac aortic dilatation imaging data: among a large number of collected aortic dilatation MRI images, convert the image sequence into NIFTI format according to the image details; and unify the image selection into instance 2 data; Obtain cardiac long-axis image data: among a large number of cardiac long-axis MRI images collected, convert the image sequence into NIFTI format according to the image details; in the selection of long-axis NIFTI images, select 2ch images; Obtain cardiac short-axis image data: among a large number of cardiac short-axis MRI images collected from subjects, convert the image sequence into NIFTI format according to the image details; unify the image selection into instance 2 data; Obtain metabolomics data: First, the metabolomics data of a large number of subjects are preprocessed, including quality control and normalization. Then, the Cox proportional hazard model is used to study the longitudinal association between metabolites and cardiovascular disease. The importance of these metabolites is quantified using machine learning methods, and then the metabolic risk score of each sample is calculated. Obtain genomic data: First, collect summary statistics of relevant cardiovascular diseases, then use PLINK software to perform standardized quality control on the genomic data of a large number of subjects to eliminate unqualified samples and SNPs, and finally use PRSice-2 software to calculate the polygenic risk score PGS of each sample. The calculation formula of PGS is: Where i represents the i-th individual, j is the j-th SNP, and w j is the weight of the SNP, a ij is the allele copy number of the jth SNP in the ith individual.

5. The method according to claim 1, characterized in that In step S21, the ResNet model is used to filter out relevant cardiovascular features from the image data to establish an image model, which specifically includes the following steps: First, the original feature image data was converted into the input format of the ResNet model. The medical image processing software SimpleITK was used to extract the 3D slices and convert them into 2D images. The images were then normalized and resized to fit the input format of the model. Then, the ResNet model is used to extract features from the three-dimensional image of the heart. The model is used to extract feature vectors from the three three-dimensional views of the heart, and the feature vectors of the three views are spliced ​​or superimposed to form a joint feature representation of the entire heart. The principal component analysis method PCA is used to reduce the dimension of the extracted features and reduce unimportant features. Finally, the fully connected layer is removed, and the image features are processed by the projection layer and mapped to the final embedding space.

6. The method according to claim 5, characterized in that The network structure of the image model includes an input layer of ResNet50, 16 residual blocks, and a pooling layer, as follows: The first input layer: The input image is passed through a 7×7 convolutional layer to extract preliminary features, and a maximum pooling layer is used for spatial downsampling to reduce the size of the feature map; The second residual layer: contains 3 residual blocks, each block contains two convolutional layers, each convolutional layer is followed by batch normalization and ReLU activation function; The third residual layer: contains 4 residual blocks, each of which contains two convolutional layers; The fourth residual layer: contains 6 residual blocks, each of which contains two convolutional layers; The fifth residual layer: contains 3 residual blocks, each of which contains two convolutional layers; The sixth average pooling layer: pools the feature map of each channel into a single value to obtain a 2048-dimensional vector.

7. The method according to claim 1, characterized in that In step S22, the genomic data is processed by a neural network model through a fully connected layer, normalization, and activation function, and the feature results are output to the embedding space.

8. The method according to claim 1, characterized in that In step S23, the metabolomics data is processed by a neural network model through a fully connected layer, normalization, and activation function, and the feature results are output to the embedding space.

9. The method according to claim 1, characterized in that The method further comprises: Step S4: Calculate the SHAP value, analyze the risk characteristic status of the person to be investigated and visualize it, specifically: After making a prediction, the cardiovascular disease risk prediction model uses the SHAP method to interpret it. During this process, a SHAP model interpreter is constructed, and the average value of the training data is used as the background value; the SHAP value of each sample is calculated, and the impact of the features of different view images on the model prediction results is analyzed, and their contribution values ​​are evaluated; finally, a SHAP explanation diagram is generated to intuitively show how the features of the original feature imaging data, genomics data, and metabolomics data affect the model's prediction of cardiac health risks.

10. A cardiovascular disease detection system integrating medical imaging and molecular omics features, characterized in that The system comprises at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs or instructions; The processor is used to obtain and execute a computer program or instruction in a memory so that the system executes the method according to any one of claims 1 to 9.

Images