Device for multi-task collaborative assessment of obstructive coronary artery disease risk level

Abstract

This invention provides a device for assessing the risk level of obstructive coronary artery disease based on multi-task collaborative methods, relating to the field of obstructive coronary artery disease risk level assessment technology. The device includes an image acquisition module for acquiring CTA image data; a coronary artery plaque segmentation module for identifying coronary arteries and segmenting coronary artery plaques to obtain a coronary artery plaque feature map; an epicardial adipose tissue segmentation module for identifying and segmenting the heart contour to extract epicardial adipose tissue and obtain an epicardial adipose tissue feature map; a collaborative learning module for fusing the coronary artery plaque feature map and the epicardial adipose tissue feature map to obtain a fused disease risk level assessment feature map; and a disease risk level assessment module for assessing the risk level of obstructive coronary artery disease. This device utilizes a collaborative learning method to solve the problem of different feature fusions and mutual constraints, improving the ability to accurately extract features and thus improving the accuracy of disease risk assessment.

Description

Technical Field

[0001] This invention relates to the field of risk assessment technology for obstructive coronary artery disease, and more particularly to a device for assessing the risk level of obstructive coronary artery disease based on multi-task collaborative assessment. Background Technology

[0002] Cardiovascular disease has become one of the leading causes of death worldwide. Myocardial ischemia is primarily caused by epicardial coronary artery obstruction. This is due to the formation and rupture of coronary plaques, leading to blockage of the coronary arteries and insufficient blood supply to the heart. Furthermore, studies have shown a positive correlation between epicardial fat thickness and the severity of coronary artery stenosis. However, AI-based studies on obstructive coronary arteries mainly focus on quantifying coronary artery luminal stenosis, without comprehensively considering the influence of multiple factors on obstructive coronary artery disease. Therefore, the accuracy of accurately assessing the risk level of obstructive coronary artery disease is relatively low.

[0003] Currently, obstructive coronary artery disease is classified into coronary artery grades based on the degree of stenosis of the coronary artery lumen. However, this classification using only a single factor has limitations. For example, epicardial adipose tissue is a metabolically active fatty tissue within the pericardium that directly interacts with the coronary arteries. Epicardial adipose tissue has been proven to be an independent cardiovascular disease risk factor. Measuring the volume of epicardial adipose tissue can assess the severity and progression of obstructive coronary artery disease, serving as an important auxiliary means for determining further treatment. Segmenting epicardial adipose tissue aims to improve the accuracy of assessing the risk level of obstructive coronary artery disease. Patent CN110378868A provides a pericardial segmentation method based on CT sequences and a method for roughly locating the pericardial region from central slices of the pericardial sequence. This method can eliminate some interfering tissue regions and improve the speed of pericardial segmentation, but it involves many steps and a large computational load. The master's thesis "Pericardial Segmentation of CT Images Based on Convolutional Neural Networks" focuses on three-dimensional convolutional neural networks, designing and implementing a pericardial segmentation algorithm based on the three-dimensional convolutional neural network SDV-Net. While this 3D method considers the spatial information of the pericardium, the small size of the model input data leads to the loss of some information, affecting feature extraction. Patent CN115049608A provides a fully automated epicardial adipose tissue extraction system based on YOLO-V5 and U-Net. This method first uses the YOLO-V5 model to determine the spatial location of the heart, then uses the U-Net model to segment the heart contour, and finally extracts epicardial adipose tissue through thresholding. This simplifies the computational burden of epicardial adipose tissue segmentation using U-Net, but it does not consider the U-Net model's lack of ability to model long-term dependencies, leading to discontinuous and inaccurate segmentation results. Furthermore, it lacks post-processing of the pericardial segmentation results, including filling internal holes and refining edges, further affecting the accuracy of epicardial adipose tissue segmentation due to inaccurate pericardial segmentation. Similarly, calcified plaques can detach and obstruct the lumen, leading to stenosis or even occlusion. Coronary artery plaque analysis plays a role in risk assessment of obstructive coronary artery disease; however, coronary artery plaque segmentation still has some shortcomings, with relatively low accuracy. The invention patent with patent number CN115423757A provides a method and device for coronary plaque segmentation. By generating a centerline distance heatmap, it assigns a larger optimization weight to the inner region of the plaque, thereby improving the segmentation accuracy.However, this method requires coronary artery mask extraction and erosion refinement, which is easily affected by noise, gray-level inhomogeneity, and other factors, potentially impacting the accuracy of centerline extraction. Patent CN113538471A uses a pre-set plaque detection network to detect coronary artery plaques in various reference image blocks within the target coronary artery, obtaining candidate reference image blocks. These candidate blocks are then input into a pre-set plaque segmentation network to obtain the coronary artery plaque segmentation result for the candidate reference image blocks. However, using multiple networks may increase network complexity and potentially lead to overfitting and underfitting. Patent CN109087708A achieves automated plaque segmentation on the vessel wall and improves the accuracy of plaque segmentation by iteratively fine-tuning a deep learning network model, thereby increasing plaque segmentation efficiency. However, this method does not perform small-scale plaque detection and directly segments plaques; its effectiveness in handling small targets may be limited, requiring extensive data training. Patent CN113077441A provides a method for segmenting coronary artery calcified plaques. It first acquires a first medical image of the heart region, then segments the plaques into three categories using a segmentation model. However, directly using the heart region still suffers from foreground-background imbalance because the heart region includes many tissues that do not need to be segmented. Patent CN114943699A acquires non-contrast-enhanced ECG-gated integrated computed tomography (CSCT) images of calcified plaques including the heart region, and trains a segmentation model based on position-encoded images to achieve accurate segmentation of calcified plaque regions in multiple coronary artery branches. However, it lacks a post-processing step, which may lead to noise or discontinuities in some segmentation results, affecting the accuracy of the segmentation and consequently the risk level of obstructive coronary artery disease. Patent CN113033626A provides an image classification method based on multi-task collaborative learning. First, several input image classification datasets are preprocessed with dimensionality reduction, and random samples are selected to form training data. Then, an image classification optimization model with novel low-rank constraints is constructed, approximating the rank minimization problem by minimizing singular values, and an iterative algorithm is optimized to solve the model. Finally, the learned mapping matrix is ​​used to classify the image dataset. However, this data iteration method is time-consuming and prone to getting trapped in local minima. Patent CN113711317A determines the pattern of functional coronary artery disease based on the blood flow mechanism of coronary arteries, but it does not consider the impact of plaque and epicardial fat on coronary artery disease and does not assess the risk level of obstructive coronary artery disease. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a device for assessing the risk level of obstructive coronary artery disease based on multi-task collaborative assessment, thereby realizing the assessment of the risk level of obstructive coronary artery disease.

[0005] To address the aforementioned technical problems, the present invention provides a device for assessing the risk level of obstructive coronary artery disease based on multi-task collaborative methods. This device includes an image acquisition module, a coronary artery plaque segmentation module, an epicardial adipose tissue segmentation module, a collaborative learning module, and a disease risk level assessment module. The image acquisition module acquires coronary CTA image data from a patient and converts the DICOM format image data to nii.gz format. The coronary artery plaque segmentation module receives the patient's CTA image data, identifies the coronary arteries, and segments the coronary artery plaques to obtain a coronary artery plaque feature map. The epicardial adipose tissue segmentation module receives the patient's CTA image data, identifies and segments the cardiac contour, and extracts epicardial adipose tissue using a thresholding method to obtain an epicardial adipose tissue feature map. The collaborative learning module fuses the coronary artery plaque feature map and the epicardial adipose tissue feature map to obtain a fused disease risk level assessment feature map. The disease risk level assessment module assesses the risk level of obstructive coronary artery disease based on the disease risk level assessment feature map.

[0006] Preferably, the coronary artery plaque segmentation module uses a region growing and region merging algorithm based on automatic threshold selection of seed points to segment the patient's coronary artery CTA image data to obtain the coronary artery region; then, morphological operations, including opening and closing operations, are performed on the coronary artery region to remove isolated pixels within the region and fill in the voids inside the plaque; a coronary artery mask image is obtained through a dilation operation; next, the coronary artery mask image is input into a U-Net network with an attention module to capture complex plaque morphology and segment the coronary artery plaque; finally, the segmentation results of the coronary artery plaque are post-processed to obtain a more refined plaque segmentation edge.

[0007] Preferably, the epicardial adipose tissue segmentation module first constructs a hybrid SwingTransformer and U-Net model and receives patient CTA image data, fusing SwingTransformer and U-Net predictions to obtain a pericardial feature map; then, it receives the pericardial feature map through a nonlocal module, capturing the interaction between all pixels to better understand the overall context information, and then outputs the pericardial feature map to the decoder; the decoder predicts the initial pericardial segmentation result using deconvolution; next, it applies morphological reconstruction opening and closing operations to the initial pericardial segmentation result to fill some holes and refine the edges, obtaining a refined pericardial segmentation result; finally, it segments the epicardial adipose tissue using a thresholding method on the patient CTA image data corresponding to the refined pericardial segmentation result to obtain an epicardial adipose tissue feature map.

[0008] Preferably, the collaborative learning module simultaneously receives coronary artery plaque feature maps and epicardial fat feature maps through two input layer paths and outputs them to a collaborative convolutional layer; after fusing the coronary artery plaque features and epicardial fat tissue features through the collaborative convolutional layer, it outputs them to a multi-layer convolutional unit; finally, the multi-layer convolutional unit further extracts the coronary artery plaque and epicardial fat tissue features to obtain a disease risk level assessment feature map.

[0009] Preferably, the collaborative convolutional layer uses convolutional units with a kernel of 3 to extract features of coronary artery plaques and epicardial fat. Then, it uses group normalization to extract features from small targets of coronary artery plaques and epicardial fat, and finally uses a nonlinear activation function to correct each feature value.

[0010] Preferably, the collaborative convolutional layer uses collaborative cross-fusion to enable mutual constraint and learning between coronary artery plaque features and epicardial adipose tissue features. Specifically, a new coronary artery plaque feature map is obtained by superimposing epicardial adipose tissue feature values ​​onto a feature map with coronary artery plaque features. Similarly, a new epicardial adipose tissue feature map is obtained by superimposing coronary artery plaque feature values ​​onto a feature map with epicardial adipose tissue features. The superposition fusion is used to fuse the above new feature maps into a feature map that simultaneously contains epicardial adipose tissue and coronary artery plaque features, which is then output to a convolutional unit with a kernel of 3 for feature extraction.

[0011] Preferably, the disease risk level assessment module receives a high-dimensional feature map of disease risk level assessment through a global average pooling unit, then maps the high-dimensional feature map of disease risk level feature map into a one-dimensional feature vector that classifies the risk level of the disease through a fully connected layer, and finally maps the above one-dimensional feature vector into a probability map between 0 and 1 through a Softmax logic layer, and determines the risk level of obstructive coronary artery disease as the maximum probability value obtained.

[0012] The beneficial effects of adopting the above technical solution are as follows: The device for assessing the risk level of obstructive coronary artery disease based on multi-task collaborative evaluation provided by the present invention improves the automated selection of seed points during the segmentation process of the region growing algorithm; overcomes the limitation of traditional region growing methods that rely entirely on manual selection of seed points; solves the problem of accurately locating and identifying small targets when the foreground and background are unbalanced; overcomes the defects of insufficient segmentation ability and lack of ability to model long-term dependencies when there are large differences between samples, resulting in discontinuous and inaccurate segmentation results; and solves the problem of different feature fusion and mutual constraints by using collaborative learning methods, improving the ability to accurately extract features, thereby improving the accuracy of disease risk assessment. Attached Figure Description

[0013] Figure 1 A structural block diagram of a device for assessing the risk level of obstructive coronary artery disease based on multi-task collaborative assessment provided in an embodiment of the present invention;

[0014] Figure 2 This is a flowchart illustrating the multi-task collaborative assessment of the risk level of obstructive coronary artery disease according to an embodiment of the present invention;

[0015] Figure 3 This is a flowchart of coronary artery plaque segmentation provided according to an embodiment of the present invention;

[0016] Figure 4 The image shows the result of coronary artery plaque segmentation according to an embodiment of the present invention, wherein (a) is a CTA image, (b) is the coronary artery segmentation result, and (c) is the plaque segmentation result;

[0017] Figure 5 A flowchart illustrating the refined segmentation of epicardial adipose tissue according to an embodiment of the present invention;

[0018] Figure 6 The image shows the result of refined epicardial adipose tissue segmentation according to an embodiment of the present invention, wherein (a) is a CTA image, (b) is a pericardial segmentation label, (c) is the pericardial segmentation result, (d) is an epicardial fat label, and (e) is the epicardial fat segmentation result.

[0019] Figure 7 A flowchart illustrating collaborative learning features provided according to embodiments of the present invention;

[0020] Figure 8 A flowchart for assessing disease risk levels provided according to embodiments of the present invention; Detailed Implementation

[0021] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0022] In this embodiment, a device for assessing the risk level of obstructive coronary artery disease is built using the Python-based PyTorch framework.

[0023] In this embodiment, the device for assessing the risk level of obstructive coronary artery disease based on multi-task collaborative evaluation, such as... Figure 1 , 2 As shown, the system includes an image acquisition module, a coronary artery plaque segmentation module, an epicardial adipose tissue segmentation module, a collaborative learning module, and a disease risk assessment module. The image acquisition module acquires the patient's coronary CTA image data and converts the DICOM format image data to nii.gz format. The coronary artery plaque segmentation module receives the patient's CTA image data, identifies the coronary arteries, and segments the coronary artery plaques to obtain a coronary artery plaque feature map. The epicardial adipose tissue segmentation module receives the patient's CTA image data, identifies and segments the cardiac contour, and extracts epicardial adipose tissue using a thresholding method to obtain an epicardial adipose tissue feature map. The collaborative learning module fuses the coronary artery plaque feature map and the epicardial adipose tissue feature map to obtain a fused disease risk assessment feature map. The disease risk assessment module assesses the risk level of obstructive coronary artery disease based on the disease risk assessment feature map.

[0024] In this embodiment, the coronary artery plaque segmentation module, such as Figure 3 As shown, region growing and region merging algorithms based on automatic threshold selection of seed points are used to segment patient coronary artery CTA image data to obtain coronary artery regions. In the region growing algorithm based on automatic threshold selection of seed points, the coronary artery segmentation result is obtained according to the seed points selected by automatic threshold and the setting of appropriate growth criteria. At the same time, the region merging algorithm is used to merge similar pixel values ​​to achieve accurate segmentation of the coronary arteries. Then, morphological operations, including opening and closing operations, are performed on the coronary artery regions to remove isolated pixels within the regions and fill the voids inside the plaques. A coronary artery mask image is obtained through dilation operation. Then, the coronary artery mask image is input into a U-Net network with an attention module to capture complex plaque morphology and segment the coronary artery plaques. Finally, the segmentation results of the coronary artery plaques are post-processed, including small region removal and smoothing, to obtain more refined plaque segmentation edges.

[0025] In this embodiment, the region growing algorithm based on automatic threshold selection of seed points includes a seed point selection strategy and a region growing criterion. The seed point selection strategy is as follows: First, the image is binarized using the existing Otsu thresholding method to obtain initial segmented regions; then, the centroid of each region in the binarized image is extracted as a candidate seed point set {p}; finally, based on the morphological features and size of the coronary artery, the most suitable seed point p0 is selected from the candidate seed points. The centroid coordinates (x, y) of the region are calculated as follows: The region growing criterion is as follows: For a seed point p0, if its neighboring pixel set is N(p0), then the condition for the current pixel p and the seed point p0 to belong to the same region is: |f(p)-f(p0)|≤T1, where f(p) represents the grayscale value or color value of pixel p, and T1 is a threshold parameter used to limit the range of similarity measurement. The region merging criterion is as follows: For two adjacent regions R1 and R2, their similarity measurement is: Where w(R) represents the weight of region R, and w(R1,R2) represents the weight between regions R1 and R2, typically using the average of the grayscale values ​​or color differences between regions as the weight. When the similarity measure S(R1,R2) is greater than the threshold T2, regions R1 and R2 are merged into one region. Dilation and erosion in morphological operations can be expressed by the following formula: Dilation: Corrosion: (A⊙B)(i,j)=min (k,l)∈B A(ik,jl), where A represents the image to be operated on, and B represents the structuring element. ⊙ and ⊙ represent the dilation and erosion operations, respectively. In the U-Net network with an attention module, the Dice loss function is used as an alternative to the cross-entropy loss function to further improve the accuracy of coronary plaque segmentation. Its formula can be expressed as: Where, p ij and g ij These represent the predicted pixels and the actual pixels, respectively. ∈ is a small positive number used to avoid a zero denominator. The smaller the value of the Dice loss function, the less overlap between the segmentation result and the true label. Post-processing of the segmentation results includes small region removal and smoothing to obtain finer patch segmentation edges. In post-processing, appropriate small region removal sizes and smoothing parameters need to be selected to obtain suitable segmentation edges. Specifically, morphological operations and image filtering can be used for post-processing. In morphological operations, opening and closing operations can be used to remove interfering factors such as small regions and isolated points. In image filtering, Gaussian filtering and other methods can be used to smooth segmentation edges to reduce noise and improve segmentation accuracy.

[0026] In this embodiment, the coronary artery plaque segmentation result for a specific patient's CTA image data is as follows: Figure 4 As shown.

[0027] In this embodiment, the epicardial adipose tissue segmentation module, such as... Figure 5 As shown, a hybrid Swing Transformer and U-Net model is first built to receive patient CTA image data, and the Swing Transformer and U-Net predictions are fused to obtain a pericardial feature map. Then, a non-local block (NLblock) is used to receive the pericardial feature map, capturing the interaction between all pixels to better understand the overall context information, and then the pericardial feature map is output to the decoder. The decoder predicts the initial pericardial segmentation result using deconvolution. Next, morphological reconstruction opening and closing operations are applied to the initial pericardial segmentation result to fill some holes and refine the edges, resulting in a refined pericardial segmentation result. Finally, the patient CTA image data corresponding to the refined pericardial segmentation result is segmented into epicardial fat tissue using a thresholding method to obtain an epicardial fat feature map.

[0028] In this embodiment, the loss function used in the hybrid SwinTransformer and U-Net model is represented by the following function:

[0029] Loss=a*lossdice+b*losscee+c*lossbdy

[0030] Where lossdice represents the Dice loss function, losscee represents the cross-entropy loss function, lossbdy represents the boundary loss function (i.e., the weighted sum of the three losses), and a, b, and c represent the weights of the three losses, respectively.

[0031] In this embodiment, the weights of the three losses are set to a = 0.4, b = 0.3, and c = 0.3, respectively. The learning rate used in the training process is set to 0.0001, decaying by a factor of 10 every 50 epochs, the number of iterations is set to 200, the optimizer is set to Adam, and the activation function is set to ReLU.

[0032] Finally, the patient's CTA image data corresponding to the refined pericardial segmentation results were segmented into epicardial adipose tissue using a threshold of [-190, -30HU]. Pixels within the threshold were set to 1, and the rest were set to 0. The resulting binary image of the epicardial adipose tissue region is the final epicardial adipose tissue segmentation result. Figure 6 As shown.

[0033] In this embodiment, as Figure 7As shown, the collaborative learning module simultaneously receives coronary artery plaque feature maps and epicardial fat feature maps through two input layer paths and outputs them to a collaborative convolutional layer. The first path receives the coronary artery plaque feature map and extracts plaque features through a convolutional unit. The second path receives the epicardial fat feature map, extracts epicardial fat features through a convolutional unit, and adjusts the number of image channels to be the same as the number of channels in the plaque feature map. Then, the weights of the coronary artery plaque feature map are superimposed on the epicardial fat feature map, and the weights of the epicardial fat feature map are superimposed on the plaque feature map to form a collaboratively constrained learning pattern. After obtaining the weighted feature maps, they are superimposed and fused to form a feature map containing both plaque and epicardial fat tissue features, which is then output to a multi-layer convolutional unit.

[0034] The coronary artery plaque and epicardial adipose tissue features are further extracted through multi-layer convolutional units. Specifically, the first convolutional unit compresses the fused feature map to half the resolution of the original image, and the second convolutional unit compresses the feature image to one-quarter of the resolution of the original image and outputs it to the third convolutional unit. The third convolutional unit further extracts disease risk level assessment feature maps with high-level semantic features for assessing disease risk levels.

[0035] In this embodiment, the disease risk level assessment module, such as... Figure 8 As shown, a high-dimensional feature map for disease risk level assessment is received through a global average pooling unit. Then, a fully connected layer maps the high-dimensional feature map of disease risk level into a one-dimensional feature vector that classifies the risk level of the disease. Finally, a Softmax logic layer maps the one-dimensional feature vector into a probability map between 0 and 1, and the maximum probability value is determined as the risk level for assessing obstructive coronary artery disease.

[0036] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. An apparatus for assessing a risk level of obstructive coronary artery disease based on multi-task synergy, the apparatus comprising: The application relates to a coronary artery plaque and epicardial adipose tissue collaborative learning method and device for evaluating the risk level of coronary artery disease. The collaborative learning module is used for fusing the coronary artery plaque feature map and the epicardial adipose tissue feature map to obtain a fused disease risk level evaluation feature map; and the disease risk level evaluation module is used for evaluating the risk level of the coronary artery disease according to the disease risk level evaluation feature map. The collaborative learning module simultaneously receives the coronary artery plaque feature map and the epicardial adipose tissue feature map through two input layer paths and outputs the coronary artery plaque feature map and the epicardial adipose tissue feature map to a collaborative convolution layer; the coronary artery plaque feature and the epicardial adipose tissue feature are fused through the collaborative convolution layer and then output to a multilayer convolution unit; finally, the coronary artery plaque feature and the epicardial adipose tissue feature are further extracted through the multilayer convolution unit to obtain the disease risk level evaluation feature map. The collaborative convolution layer is used for extracting the coronary artery plaque feature and the epicardial adipose tissue feature through a convolution unit with a convolution kernel of 3, extracting the coronary artery plaque feature and the epicardial adipose tissue feature through group normalization, and correcting the feature values through a nonlinear activation function.

2. The apparatus for risk stratification of obstructive coronary artery disease based on multi-task synergy evaluation according to claim 1, wherein: The coronary artery plaque segmentation module uses a region growing algorithm and a region merging algorithm based on automatic threshold selection seed points to segment the coronary artery CTA image data of a patient to obtain a coronary artery region; then, the coronary artery region is subjected to morphological operations, including opening operation and closing operation, to remove isolated pixels in the region and fill internal cavities of the plaque; the coronary artery mask image is obtained through inflation operation; then, the coronary artery mask image is input into a U-Net network with an attention module to capture complex plaque morphology and segment the coronary artery plaque; finally, the segmentation result of the coronary artery plaque is post-processed to obtain a more fine plaque segmentation edge.

3. The apparatus for risk stratification of obstructive coronary artery disease based on multi-task synergy evaluation according to claim 2, characterized in that: The epicardial adipose tissue segmentation module first builds a hybrid Swin Transformer and U-Net model, receives the CTA image data of the patient, and fuses the Swin Transformer and the U-Net to obtain a pericardium feature map; then, the pericardium feature map is input into a non-local module to capture the interaction between all pixels and better understand the whole context information, and the pericardium feature map is output to a decoder. The decoder predicts the pericardium initial segmentation result by the method of deconvolution; then the pericardium initial segmentation result is used to perform the open and close operation of morphological reconstruction, a part of hole filling and edge thinning of the pericardium segmentation result are performed, and a refined pericardium segmentation result is obtained; finally, the pericardium fat feature map is obtained by performing the threshold method on the patient CTA image data corresponding to the refined pericardium segmentation result to segment the epicardial fat tissue.

4. The apparatus for risk stratification of obstructive coronary artery disease based on multi-task synergy evaluation according to claim 3, characterized in that: The synergistic convolution layer is used for mutual restriction and learning between the coronary artery plaque features and the epicardial fat tissue features through synergistic cross fusion, specifically, the feature map with the coronary artery plaque features is superimposed with the epicardial fat tissue feature values to obtain a new coronary artery plaque feature map, and similarly, the feature map with the epicardial fat tissue features is superimposed with the coronary artery plaque feature values to obtain a new epicardial fat tissue feature map, and the above new feature maps are fused into a feature map with both epicardial fat tissue and coronary artery plaque tissue by superimposition fusion, and output to a convolution unit with a convolution kernel of 3 for feature extraction.

5. The apparatus for risk stratification of obstructive coronary artery disease based on multi-task synergy evaluation according to claim 4, characterized in that: The disease risk level evaluation module receives the disease risk level evaluation feature map through the global average pooling unit, then maps the high-dimensional features of the disease risk level evaluation feature map into a one-dimensional feature vector for dividing the risk level of the disease through the fully connected layer, and finally maps the above one-dimensional feature vector into a probability map between 0 and 1 through the Softmax logic layer, and determines the maximum probability value as the risk level of the evaluated obstructive coronary artery disease.

Images