Deep learning based ct image recognition method for large animal arterial calcification

Abstract

The present application relates to the technical field of image processing, and discloses a large animal arterial calcification CT image recognition method based on deep learning, which comprises the following steps: acquiring angiography image data, extracting an arterial vessel tree center line, and generating a vessel center line distance map; inputting the image data and the vessel center line distance map into a three-dimensional segmentation network to obtain a preliminary segmentation result and a segmentation probability map; performing connected domain analysis on the preliminary segmentation result, calculating a structure confidence score, weighting the segmentation probability map lesion by lesion using the structure confidence score, and performing threshold processing to obtain an arterial calcification region segmentation result; calculating a high-flux image feature set, performing feature clustering, selecting representative features from the clustering community to form an optimal feature subset; and inputting the optimal feature subset into a classifier to obtain a recognition result. The present application can not only accurately recognize known defect types, but also effectively detect new defects that have not been learned, greatly enhancing the generalization ability and robustness of detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a deep learning-based CT image recognition method for large animal arterial calcification. BACKGROUND

[0002] Arterial calcification is a pathological link closely related to cardiovascular event risk in the development of cardiovascular diseases. In preclinical studies, large animal models such as pigs and dogs are ideal carriers for studying the mechanisms of atherosclerosis and calcification and evaluating the effectiveness of new treatment options because their cardiovascular systems are highly similar to humans in terms of anatomy and physiology. CT angiography (CTA) is the main imaging method for observing and quantitatively evaluating the calcification load of large animal arteries in vivo. However, the calcification evaluation method is highly dependent on manual or semi-automatic delineation by radiologists or researchers, which is not only time-consuming and labor-intensive but also subjective and observer-dependent, making it difficult to meet the requirements of large-scale high-throughput research for data processing efficiency and result consistency. Deep learning models represented by U-shaped networks (U-Net) have been applied to medical image segmentation tasks, achieving automatic identification of arterial calcification regions to some extent and improving work efficiency. However, due to the diverse shapes and sizes of arterial calcification lesions and the similar signal intensity of high-density tissues such as bone and vascular stents on CT images, the segmentation network is prone to missed detection or missegmentation, making it difficult to ensure the accuracy and integrity of the segmentation results. Radiomics technology extracts a large number of quantitative features from medical images, but the high-dimensional feature set has serious information redundancy and collinearity problems. How to select the optimal feature subset with the highest correlation with disease status and information complementarity from the vast amount of features is a key bottleneck for the clinical application of radiomics. There is an urgent need for an automatic recognition method that integrates high-precision segmentation and efficient feature selection to achieve end-to-end intelligent analysis from raw CT images to large animal arterial calcification risk classification or type recognition. SUMMARY

[0003] The present application provides a deep learning-based CT image recognition method for large animal arterial calcification to solve the problem of easy missed detection or missegmentation in the prior art, resulting in poor accuracy of the recognition results.

[0004] The deep learning-based CT image recognition method for large animal arterial calcification of the present application includes the following steps:

[0005] The three-dimensional CT angiography image data of a large animal is acquired, the center line of an arterial blood vessel tree is extracted based on the image data, the distance from each voxel to the nearest center line is calculated, and a blood vessel center line distance map is generated; the three-dimensional CT angiography image data of the large animal and the blood vessel center line distance map are taken as multi-channel inputs and input to a pre-trained three-dimensional segmentation network to obtain a preliminary segmentation result of an arterial calcification region and a corresponding voxel-level segmentation probability map; the three-dimensional segmentation network is trained using a combined loss function comprising a Dice loss and a Focal Loss adjusted according to the voxel Hounsfield unit value; connected component analysis is performed on the preliminary segmentation result, topological and morphological indexes are extracted for each candidate calcification lesion, the topological and morphological indexes are input to a pre-trained Bayesian network to calculate a structure confidence score, the structure confidence score is used to weight the voxel-level segmentation probability map on a lesion-by-lesion basis, threshold processing is performed on the weighted voxel-level segmentation probability map to obtain an arterial calcification region segmentation result; within the arterial calcification region segmentation result, a high-throughput image feature set is calculated, representative features are selected from each clustering community by constructing an inter-feature mutual information matrix and using a community detection algorithm to perform feature clustering, and an optimal feature subset is formed; the optimal feature subset is input into a pre-set gradient boosting decision tree classifier to obtain a recognition result of the risk level of the arterial calcification state of the large animal.

[0006] Preferably, the combined loss function is L = a L Dice +(1-a) L weighted_focal , wherein is the Dice loss, is the weighted Focal Loss, and is a balance coefficient in the range of [0, 1]; in the weighted Focal Loss, for any voxel, when its Hounsfield unit value is greater than 130, the weight is 2.0, otherwise the weight is 0.5.

[0007] Preferably, the centerline of the arterial vessel tree is extracted based on the image data, comprising: inputting the three-dimensional CT angiography image data of the large animal into a key point detection deep learning network, identifying and outputting three-dimensional coordinates of a centerline starting point and a centerline ending point; the centerline starting point corresponds to the root of the arterial vessel tree to be analyzed, and the centerline ending point corresponds to the distal end of the branch of the arterial vessel tree; applying a vessel enhancement filter to the three-dimensional CT angiography image data of the large animal to generate a vessel response map, and taking the vessel response map as a speed function of a fast marching algorithm; taking the detected centerline starting point as a source point, propagating on the speed map by using the fast marching algorithm, calculating the minimum arrival time from the source point to each voxel in the three-dimensional CT angiography image of the large animal, and generating an arrival time map; respectively starting from each detected centerline ending point, backtracking in the gradient descent direction on the arrival time map until the centerline starting point is reached, and the path sequence formed constitutes the centerline of the arterial vessel tree.

[0008] Preferably, the topological and morphological indicators include volume, surface area, sphericity, and maximum three-dimensional diameter.

[0009] Preferably, the voxel-level segmentation probability map is weighted by the structure confidence score on a lesion-by-lesion basis, comprising: multiplying the structure confidence score of each candidate calcification lesion by the original probability value of each voxel in the candidate calcification lesion to obtain a modulated probability value.

[0010] Preferably, the feature clustering comprises: using the Louvain community detection algorithm to process the mutual information matrix between the high-throughput image features, and dividing the high-throughput image features into different communities.

[0011] Preferably, the representative feature is selected from each clustered community, comprising: within each community, calculating the sum of the mutual information values of a single high-throughput image feature with all other high-throughput image features in the community, and selecting the high-throughput image feature with the largest sum value as the representative feature of the community.

[0012] Preferably, the risk level of the arterial calcification state of the large animal is one of low risk, medium risk and high risk.

[0013] Preferably, the Bayesian network is a tree augmented naive Bayesian network.

[0014] Preferably, the three-dimensional segmentation network is an nnU-Net segmentation network.

[0015] The beneficial effects of the present application are: the present application takes the blood vessel center line distance map as the network input, adopts a combined loss function with the weight adjusted according to the Hounsfield unit value, enhances the ability of the segmentation network to distinguish the calcified lesions in the artery from the high-density tissues outside the blood vessel such as bones, and improves the segmentation accuracy of small and irregularly shaped calcified lesions. Moreover, the preliminary segmentation result is weighted and corrected by using the topological and morphological prior knowledge of the lesion, so as to eliminate the false positive regions caused by artifacts and the like, and ensure the high reliability of the segmentation result. In addition, in the imaging omics analysis stage, the mutual information matrix is constructed, and the feature screening is performed by combining the community detection algorithm, so that the optimal feature subset with large information amount and low redundancy can be efficiently selected from the mass of high-dimensional features. Moreover, when the arterial vessel tree center line is extracted, the starting point and the ending point are automatically realized by using the key point detection deep learning network. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A flowchart of a deep learning-based large animal arterial calcification CT image recognition method provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0017] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings. The embodiments described below by reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0018] As shown in Figure 1 The deep learning-based large animal arterial calcification CT image recognition method provided by the embodiment of the present application specifically includes the following steps:

[0019] S1, acquiring large animal three-dimensional CT angiography image data, extracting the arterial vessel tree center line based on the image data, calculating the distance from each voxel to the nearest center line, and generating a blood vessel center line distance map.

[0020] The original large animal three-dimensional CT angiography image data is processed by using a blood vessel modeling toolkit VMTK, the main aorta and main branch blood vessels are segmented by using a level set method, and the center line coordinate point set of the arterial vessel tree is extracted based on the blood vessel segmentation result by using a Voronoi graph algorithm. For each voxel in the large animal three-dimensional CT angiography image data, the Euclidean distance between the three-dimensional coordinates of the voxel and all points in the center line point set is calculated, and the minimum value is taken as the distance value of the voxel. All distance values of the voxels constitute a blood vessel center line distance map with the same size as the original large animal three-dimensional CT angiography image.

[0021] Extracting the centerline of the arterial tree can accurately locate and analyze the position of each calcified lesion in the arterial tree. In this embodiment, the fast marching algorithm is used to extract the centerline of the arterial tree. The fast marching algorithm performs well in terms of computational efficiency and accuracy, and is suitable for processing complex vascular bifurcation structure. The fast marching algorithm simulates the process of wave front propagation from one or more starting points in the vascular lumen, and can robustly generate smooth and centered centerline.

[0022] In actual operation, a starting point is manually placed on the starting layer of the big animal three-dimensional CT angiography image data sequence, for example, a starting point is manually placed on the root of the ascending aorta. The fast marching algorithm takes the starting point as the source of the wave front, and takes the gray value of the voxel as the speed function. The wave front propagates faster in the high gray value area of the vascular lumen and slower in the low gray value area of the vascular wall or outside. The algorithm calculates the shortest time for the wave front to reach each voxel in the blood vessel, and generates a time of arrival map. By backtracking the gradient descent path in the time of arrival map, the shortest time path from any end point of the arterial tree to the starting point can be traced back, which constitutes the centerline of the corresponding vascular branch. For example, backtracking from the end of the iliac artery to the root of the aorta can accurately extract the centerline of the abdominal aorta to the iliac artery segment.

[0023] In order to avoid the problem of inaccurate and low efficiency of manual selection of seed points, in an optional embodiment, the centerline of the arterial tree is extracted based on the image data, comprising: inputting the big animal three-dimensional CT angiography image data into a key point detection deep learning network to identify and output the three-dimensional coordinates of the centerline starting point and the centerline end point; the centerline starting point corresponds to the root of the arterial tree to be analyzed, and the centerline end point corresponds to the end of the branch of the arterial tree; applying a blood vessel enhancement filter to the big animal three-dimensional CT angiography image data to generate a blood vessel response map, and taking the blood vessel response map as the speed function of the fast marching algorithm; taking the detected centerline starting point as the source point, propagating on the speed map using the fast marching algorithm, calculating the minimum arrival time from the source point to each voxel in the big animal three-dimensional CT angiography image, and generating an arrival time map; starting from each detected centerline end point, respectively, backtracking in the gradient descent direction on the arrival time map until the centerline starting point is reached, and the path sequence formed constitutes the centerline of the arterial tree.

[0024] S2, input the big animal three-dimensional CT angiography image data and the blood vessel centerline distance map as a multi-channel input into a pre-trained three-dimensional segmentation network to obtain a preliminary segmentation result of the arterial calcification region and a corresponding voxel-level segmentation probability map; the three-dimensional segmentation network is trained using a combined loss function comprising Dice loss and Focal Loss with weight adjusted according to the voxel Hounsfield unit value.

[0025] The three-dimensional CT angiography image data of a large animal is taken as a first channel, and the blood vessel centerline distance map is taken as a second channel, which are stacked in the channel dimension to form a two-channel three-dimensional image tensor. The three-dimensional image tensor is input into a pre-trained three-dimensional segmentation network. The three-dimensional segmentation network can adopt an nnU-Net segmentation network. The combined loss function is a linear weighted sum of the Dice loss and the weighted Focal Loss. The weight of the Focal Loss is set according to the Hounsfield unit value of the voxel, for example, the voxel with a Hounsfield unit value higher than 130 is assigned a weight of 1.0, and the voxel with a Hounsfield unit value lower than 130 is assigned a weight of 0.25, so as to guide the network to pay more attention to the high-density area.

[0026] In an optional embodiment, the combined loss function is defined as L = a L Dice + (1-a) L weighted_focal , where is the Dice loss, is the weighted Focal Loss, and is a balance coefficient in the range of [0, 1]. In the weighted Focal Loss, for any voxel, when its Hounsfield unit value is greater than 130, the weight is 2.0, otherwise the weight is 0.5.

[0027] The combined loss function combines the advantages of the Dice loss and the weighted Focal Loss. The Dice loss can effectively handle the serious imbalance between positive and negative samples, which is common in the calcification lesion segmentation task, because the calcification area is very small compared to the entire scan image. The Focal Loss focuses on the model's attention to the voxels that are difficult to classify, i.e., the real calcification points with low prediction probability or the non-calcification points with high prediction probability. By adjusting the contribution of the two through the balance coefficient , for example, when is set to 0.6, it means that the model training focuses more on optimizing the overlap of the segmented area.

[0028] The arterial calcified plaque is a high-density tissue, and its Hounsfield unit value is usually much higher than that of the surrounding soft tissue and blood, generally exceeding 130. Therefore, a high weight of 2.0 is assigned to the voxel with a Hounsfield unit value greater than 130, so that the model focuses on learning the area that is most likely to be a calcification lesion. Conversely, for voxels with a Hounsfield unit value not greater than 130, such as blood vessel cavities, soft tissue or image noise, a low weight of 0.5 is assigned to reduce their interference with model training. This makes the model not only learn the context features of the image, but also directly use the physical density information of the voxel, reducing the misjudgment of high-density artifacts.

[0029] S3, performing connected component analysis on the preliminary segmentation result, extracting topological and morphological indexes for each candidate calcification lesion, inputting the topological and morphological indexes into a pre-trained Bayesian network to calculate a structure confidence score, performing lesion-by-lesion weighting on the voxel-level segmentation probability map using the structure confidence score, performing threshold processing on the weighted voxel-level segmentation probability map to obtain an arterial calcification region segmentation result.

[0030] Specifically, a three-dimensional 26 connected component algorithm is used to analyze the preliminary segmentation result to identify each candidate calcification lesion. For each connected component, more than ten topological and morphological indexes such as the volume, surface area, sphericity, elongation, and average centerline distance of the connected component are extracted. These indexes are input as nodes into a pre-trained tree-augmented naive Bayesian network, and the tree-augmented naive Bayesian network outputs a structure confidence score between 0 and 1. The voxel probability values in the voxel-level segmentation probability map corresponding to each connected component are multiplied by the confidence score of the corresponding connected component to obtain a weighted voxel-level segmentation probability map. Threshold processing is performed on the weighted voxel-level segmentation probability map to generate an arterial calcification region segmentation mask, where the threshold is 0.5.

[0031] In an optional embodiment, the topological and morphological indexes include: volume, surface area, sphericity, and maximum three-dimensional diameter.

[0032] After all potential calcification lesions are preliminarily segmented, in order to quantitatively analyze and distinguish them, a set of key topological and morphological features are extracted for each connected lesion region. These features can describe the size, shape, and complexity of the lesion. Taking a candidate lesion as an example, the feature extraction process is as follows: the volume is calculated by accumulating the number of all voxels constituting the lesion and multiplying by the physical volume of a single voxel; the surface area is obtained by calculating the total area of the exposed surface of the lesion boundary voxels; the sphericity is calculated based on the volume and surface area; and the maximum three-dimensional diameter refers to the maximum Euclidean distance between any two points on the surface of the lesion.

[0033] In an optional embodiment, the lesion-by-lesion weighting of the voxel-level segmentation probability map using the structure confidence score includes: multiplying the structure confidence score of each candidate calcification lesion by the original probability value of each voxel in the candidate calcification lesion to obtain a modulated probability value.

[0034] The voxel-level segmentation probability map only reflects the local possibility of each voxel belonging to calcified tissue, but does not consider whether the entire connected region to which the voxel belongs has the typical shape of a true calcification. By introducing a structure confidence score, the structure confidence score comprehensively evaluates the overall shape, size and texture of the candidate lesion, providing a holistic judgment of the authenticity of the entire lesion. For example: suppose two candidate lesions are segmented, lesion A and lesion B. Lesion A has a regular shape, a medium size, and a structure confidence score of 0.95, indicating that lesion A is very likely to be a true calcification. The original probability value of a voxel inside lesion A is 0.8, and after modulation, the probability value of the voxel becomes 0.76. Lesion B has an elongated shape and a fuzzy edge, which may be caused by imaging artifacts, and the structure confidence score is only 0.2. The original probability value of a voxel inside lesion B is very high, such as 0.85, and the modulated probability value is greatly suppressed to 0.17, which can be effectively filtered out in the subsequent threshold processing, thereby preventing misjudgment.

[0035] S4, in the arterial calcification region segmentation result, a high-throughput image feature set is calculated, feature clustering is performed by constructing an inter-information matrix between the features and using a community detection algorithm, representative features are selected from each clustering community to form an optimal feature subset; the optimal feature subset is input into a pre-set gradient boosting decision tree classifier to obtain a recognition result of the risk level of the arterial calcification state of the large animal.

[0036] The high-throughput image feature set is also called an imageomics feature, which not only includes basic morphological indicators such as the volume and shape of the lesion, but also contains hundreds of refined indicators for describing the internal pixel intensity distribution, spatial relationship and complex texture pattern. Using the PyRadiomics toolkit, more than one hundred high-throughput image features are extracted from the three-dimensional CT angiography image data of the large animal within the region defined by the arterial calcification region segmentation mask, including first-order statistical features, gray level co-occurrence matrix features, gray level run length matrix features and gray level zone size matrix features. The normalized mutual information between all features is calculated to construct a feature inter-information matrix. The feature inter-information matrix is regarded as the adjacency matrix of a weighted undirected graph, and the Louvain community detection algorithm is used to cluster the high-throughput image features. For each detected community, select the feature with the maximum sum of mutual information values with all other features in the community as the representative feature of the community, and all representative features together constitute an optimal feature subset. The optimal feature subset vector composed of representative features is input into an XGBoost gradient boosting decision tree classifier trained in advance using labeled data. The classifier outputs the recognition result of the arterial calcification state according to the input feature vector.

[0037] In an optional embodiment, the feature clustering comprises: processing the mutual information matrix among the high-throughput image features by using a Louvain community detection algorithm, and dividing the high-throughput image feature clusters into different communities.

[0038] When hundreds of high-throughput image features are extracted from calcified lesions, the problem of high dimensionality and information redundancy of features will be encountered. In order to solve this problem, a feature clustering technique is used to group functionally or informationally similar features. Mutual information is used as a measure of the correlation between features. Specifically, the mutual information value between all pairs of features is calculated to form an N by N mutual information matrix, where N is the total number of features, such as 500. Each element value in the matrix represents the amount of information shared by the corresponding two features. This matrix is regarded as a network graph, where the features are nodes and the mutual information values are edge weights. The Louvain community detection algorithm is applied to this network graph, and the features are divided into different communities by optimizing the modularity index. For example, one community gathers all the features related to lesion volume and size, another community gathers the features related to gray level histogram statistics, and a community contains the features describing the sharpness of the lesion edge.

[0039] In an optional embodiment, the representative features are selected from each clustering community, comprising: calculating the sum of mutual information values of a single high-throughput image feature with all other high-throughput image features in the community within each community, and selecting the high-throughput image feature with the largest sum value as the representative feature of the community.

[0040] In an optional embodiment, the identification result of the risk level of the calcification state of the large animal artery is one of low risk, medium risk and high risk.

[0041] Low risk can correspond to a total calcification volume less than 10 cubic millimeters and no large plaque, representing early disease or no obvious lesion. Medium risk can correspond to a calcification volume between 10 and 100 cubic millimeters, with lesions scattered, representing that the disease has formed but has not yet reached the most severe degree. High risk can correspond to a total calcification volume greater than 100 cubic millimeters, or the presence of a fused large plaque with a length of more than 10 millimeters.

[0042] The implementation principle of the deep learning-based large animal arterial calcification CT image recognition method in the embodiment of the application is as follows: the application takes the blood vessel center line distance map as the network input, adopts a combined loss function adjusted according to the Hounsfield unit value, can enhance the segmentation network's ability to distinguish the calcification lesions in the artery from the high-density tissues outside the blood vessel such as bones, and improves the segmentation accuracy of small and irregularly shaped calcification lesions. In addition, the confidence weighting and correction of the preliminary segmentation result by using the topological and morphological prior knowledge of the lesion can eliminate the false positive regions caused by artifacts and the like, thereby ensuring the high reliability of the segmentation result. Moreover, by constructing a mutual information matrix and combining a community detection algorithm for feature screening, the optimal feature subset with large information amount and low redundancy can be efficiently selected from the mass of high-dimensional features. In addition, when extracting the arterial vessel tree center line, the starting point and the ending point are automatically realized by using a key point detection deep learning network.

[0043] Although the embodiments of the application have been shown and described above, it should be understood that the above-described embodiments are exemplary, and should not be construed as limiting the application, and those of ordinary skill in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the application.

Claims

1. A CT image recognition method for large animal arterial calcification based on deep learning, characterized in that, The method comprises the following steps: acquiring three-dimensional CT angiography image data of a large animal, extracting an arterial vessel tree center line based on the image data, calculating a distance from a voxel to a nearest center line for each voxel, and generating a vessel center line distance map; inputting the three-dimensional CT angiography image data of the large animal and the vessel center line distance map as a multi-channel input into a pre-trained three-dimensional segmentation network to obtain a preliminary segmentation result of an arterial calcification region and a corresponding voxel-level segmentation probability map; the three-dimensional segmentation network is trained using a combined loss function comprising a Dice loss and a Focal Loss with a weight adjusted according to a voxel Hounsfield unit value; performing connected component analysis on the preliminary segmentation result, extracting topological and morphological indexes for each candidate calcification lesion, inputting the topological and morphological indexes into a pre-trained Bayesian network to calculate a structure confidence score, weighting the voxel-level segmentation probability map lesion by lesion using the structure confidence score, and performing threshold processing on the weighted voxel-level segmentation probability map to obtain an arterial calcification region segmentation result; and calculating a high-throughput image feature set within the arterial calcification region segmentation result, clustering the high-throughput image features by constructing an inter-feature mutual information matrix and using a community detection algorithm, selecting representative features from each clustering community to form an optimal feature subset, and inputting the optimal feature subset into a pre-set gradient boosting decision tree classifier to obtain a recognition result of a risk level of the arterial calcification state of the large animal. The combined loss function is L = a L Dice +(1-a) L weighted_focal wherein is the Dice loss, is the weighted Focal Loss, is a balancing coefficient in the range [0, 1]; In the weighted Focal Loss, for any voxel, when the Hounsfield unit value is greater than 130, the weight is 2.0, otherwise the weight is 0.

5. The arterial vessel tree center line is extracted based on the image data, and the method comprises the following steps: inputting the three-dimensional CT angiography image data of the large animal into a key point detection deep learning network to identify and output three-dimensional coordinates of a center line starting point and a center line ending point; the center line starting point corresponds to a root of an arterial vessel tree to be analyzed, and the center line ending point corresponds to a distal end of a branch of the arterial vessel tree; applying a vessel enhancement filter to the three-dimensional CT angiography image data of the large animal to generate a vessel response map, and using the vessel response map as a speed function of a fast marching algorithm; using the detected center line starting point as a source point, propagating on the speed map using the fast marching algorithm to calculate a minimum arrival time from the source point to each voxel in the three-dimensional CT angiography image of the large animal, and generating an arrival time map; starting from each detected center line ending point, backtracking in the gradient descent direction on the arrival time map until the center line starting point is reached, and the path sequence formed constitutes the arterial vessel tree center line. 2.The deep learning-based method for identifying CT images of arterial calcification in large animals according to claim 1, characterized in that, The topological and morphological indexes include: volume, surface area, sphericity, and maximum three-dimensional diameter. 3.The deep learning-based method for identifying CT images of arterial calcification in large animals according to claim 1, characterized in that, The voxel-level segmentation probability map is weighted lesion by lesion using the structure confidence score, and the method comprises the following steps: multiplying the structure confidence score of each candidate calcification lesion by the original probability value of each voxel in the candidate calcification lesion to obtain a modulated probability value. 4.The deep learning-based method for identifying CT images of arterial calcification in large animals according to claim 1, characterized in that, The feature clustering comprises the following steps: using a Louvain community detection algorithm to process the inter-information matrix of the high-throughput image features, and clustering the high-throughput image features into different communities. 5.The deep learning-based method for identifying CT images of arterial calcification in large animals according to claim 4, characterized in that, The screening of the representative features from each cluster community comprises: In each community, the sum of mutual information values of a single high-throughput image feature and all other high-throughput image features in the community is calculated, and the high-throughput image feature with the maximum sum value is selected as the representative feature of the community. 6.The deep learning-based method for identifying CT images of arterial calcification in large animals according to claim 1, characterized in that, The identification result of the risk level of the arterial calcification state of the large animal is one of low risk, medium risk and high risk. 7.The deep learning-based method for identifying CT images of arterial calcification in large animals according to claim 1, characterized in that, The Bayesian network is a tree augmented naive Bayesian network. 8.The deep learning-based method for identifying CT images of arterial calcification in large animals according to claim 1, characterized in that, The three-dimensional segmentation network is an nnU-Net segmentation network.

Images