Large animal artery calcification CT image recognition method based on deep learning

Abstract

The invention relates to the technical field of image processing, and discloses a large animal artery calcification CT image recognition method based on deep learning, comprising: acquiring angiography image data, extracting an artery vessel tree center line, and generating a vessel center line distance map; inputting the image data and the blood vessel center line distance graph into a three-dimensional segmentation network to obtain a preliminary segmentation result and a segmentation probability graph; performing connected domain analysis on the preliminary segmentation result, calculating a structure confidence score, and performing focus-by-focus weighting and threshold processing on the segmentation probability graph by using the structure confidence score to obtain an artery calcification region segmentation result; a high-throughput image feature set is calculated, feature clustering is carried out, representative features are screened from clustering communities, and an optimal feature subset is formed; and inputting the optimal feature subset into a classifier to obtain an identification result. According to the method, known defect types can be accurately identified, novel defects which are not learned can be effectively detected, and the generalization ability and robustness of detection are greatly enhanced.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method for recognizing CT images of large animal arterial calcification based on deep learning. Background Technology

[0002] Arterial calcification is a pathological step closely related to the risk of cardiovascular events during the development of cardiovascular disease. In preclinical studies, large animal models such as pigs and dogs are ideal for studying the mechanisms of atherosclerosis and calcification, and evaluating the effectiveness of novel treatments, due to the high anatomical and physiological similarity of their cardiovascular systems to humans. CT angiography (CTA) is the primary imaging method for in vivo observation and quantitative assessment of arterial calcification load in large animals. However, calcification assessment methods heavily rely on manual or semi-automatic delineation by radiologists or researchers. This process is not only time-consuming and labor-intensive but also subject to subjectivity and inter-observer variability, making it difficult to meet the requirements of large-scale, high-throughput studies for data processing efficiency and result consistency. Deep learning models, represented by 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 morphologies and sizes of arterial calcification lesions, and their often similar signal intensities to high-density tissues such as bone and vascular stents on CT images, segmentation networks are prone to missed detections or missegments, making it difficult to guarantee the accuracy and completeness of segmentation results. Radiomics technology extracts a large number of quantitative features from medical images in high throughput, but the resulting high-dimensional feature sets suffer from serious information redundancy and collinearity problems. How to select the optimal feature subset that is most relevant to the disease state and complementary to the massive feature set is a key bottleneck in the clinical translation of radiomics. Currently, there is an urgent need for an automatic identification method that can integrate high-precision segmentation and efficient feature selection to achieve end-to-end intelligent analysis from raw CT images to the risk grading or type identification of arterial calcification in large animals. Summary of the Invention

[0003] This invention provides a deep learning-based method for recognizing large animal arterial calcification CT images to solve the problems of poor accuracy in the above-mentioned prior art, which is prone to missed detections or missegmentation.

[0004] The present invention provides a deep learning-based method for recognizing large animal arterial calcification CT images, comprising the following steps: Three-dimensional CT angiography images of large animals were acquired. Arterial vascular tree centerlines were extracted from the image data. The distance from each voxel to the nearest centerline was calculated, generating a vascular centerline distance map. The three-dimensional CT angiography images and the vascular centerline distance map were used as multi-channel inputs to a pre-trained three-dimensional segmentation network to obtain preliminary segmentation results of arterial calcification regions and corresponding voxel-level segmentation probability maps. The three-dimensional segmentation network employed a Focal loss algorithm that included Dice loss and weights adjusted according to the voxel Heinz unit value. The combined loss function of Loss is trained; connected component analysis is performed on the preliminary segmentation results to extract topological and morphological indicators for each candidate calcified lesion. The topological and morphological indicators are input into a pre-trained Bayesian network to calculate the structure confidence score. The structure confidence score is used to weight the voxel-level segmentation probability map lesion by lesion. Thresholding is performed on the weighted voxel-level segmentation probability map to obtain the segmentation result of the arterial calcification region; within the segmentation result of the arterial calcification region, a high-throughput image feature set is calculated. By constructing a mutual information matrix between features and using a community detection algorithm to perform feature clustering, representative features are selected from each cluster to form the optimal feature subset; the optimal feature subset is input into a preset gradient boosting decision tree classifier to obtain the identification result of the risk level of arterial calcification status in large animals.

[0005] Preferably, the combined loss function is L=α L Dice +(1-α) L weighted_focal ,in For the loss of Des, For weighted Focal Loss, The balance coefficient is in the range of [0, 1]. In the weighted Focal Loss, for any voxel, the weight is 2.0 when its Heinz unit value is greater than 130, and 0.5 otherwise.

[0006] Preferably, the extraction of the arterial vascular tree centerline based on image data includes: inputting the large animal 3D CT angiography image data into a keypoint detection deep learning network, identifying and outputting the 3D coordinates of the centerline start point and the centerline end point; the centerline start point corresponds to the root of the arterial vascular tree to be analyzed, and the centerline end point corresponds to the end of a branch of the arterial vascular tree; applying a vascular enhancement filter to the large animal 3D CT angiography image data to generate a vascular response map, and using the vascular response map as the velocity function of a fast traversal algorithm; using the detected centerline start point as the source point, propagating on the velocity map using the fast traversal algorithm, calculating the minimum arrival time from the source point to each voxel in the large animal 3D CT angiography image, and generating an arrival time map; starting from each detected centerline end point, backtracking along the gradient descent direction on the arrival time map until reaching the centerline start point, and the resulting path sequence constitutes the arterial vascular tree centerline.

[0007] Preferably, the topological and morphological parameters include: volume, surface area, sphericity, and maximum three-dimensional diameter.

[0008] Preferably, the step of using the structural confidence score to perform lesion-by-lesion weighting on the voxel-level segmentation probability map includes: multiplying the structural confidence score of each candidate calcified lesion by the original probability value of each voxel within the candidate calcified lesion to obtain the modulated probability value.

[0009] Preferably, the feature clustering includes: processing the mutual information matrix between high-throughput image features using the Louvain community detection algorithm, and dividing the high-throughput image features into different communities.

[0010] Preferably, the step of selecting representative features from each cluster community includes: within each community, calculating the sum of mutual information values ​​between a single high-throughput image feature and all other high-throughput image features within the community, and selecting the high-throughput image feature with the largest sum as the representative feature of the community.

[0011] Preferably, the risk level of the arterial calcification status in large animals is identified as one of three levels: low risk, medium risk, and high risk.

[0012] Preferably, the Bayesian network is a tree-augmented naive Bayesian network.

[0013] Preferably, the three-dimensional segmentation network is the nnU-Net segmentation network.

[0014] The beneficial effects of this invention are as follows: This invention uses a vascular centerline distance map as network input and employs a combined loss function with weights adjusted according to Henlein unit values. This enhances the segmentation network's ability to distinguish between intra-arterial calcified lesions and extra-vascular high-density tissues such as bone, improving the segmentation accuracy for small and irregularly shaped calcified lesions. Furthermore, by utilizing prior knowledge of lesion topology and morphology to perform confidence weighting and correction on the preliminary segmentation results, false positive regions caused by artifacts and other factors are eliminated, ensuring high reliability of the segmentation results. In addition, during the radiomics analysis stage, by constructing a mutual information matrix and combining it with a community detection algorithm for feature selection, the optimal feature subset with high information content and low redundancy can be efficiently selected from massive high-dimensional features. Moreover, when extracting the arterial vascular tree centerline, a keypoint detection deep learning network automates the selection of start and end points. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the deep learning-based method for recognizing large animal arterial calcification CT images, as provided in an embodiment of the present invention. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0017] like Figure 1 As shown, the deep learning-based method for recognizing large animal arterial calcification CT images provided in this embodiment of the invention specifically includes the following steps: S1. Acquire three-dimensional CT angiography image data of large animals, extract the center line of the arterial vascular tree based on the image data, calculate the distance of each voxel to the nearest center line, and generate a vascular center line distance map.

[0018] The original 3D CT angiography images of large animals were processed using the VMTK vascular modeling toolkit. The aorta and major branch vessels were segmented using the level set method. Based on the segmentation results, the Voronoi diagram algorithm was used to extract the centerline coordinate point set of the arterial vascular tree. For each voxel in the 3D CT angiography image data of the large animal, the Euclidean distance between the voxel's 3D coordinates and all points in the centerline point set was calculated. The minimum value was taken as the voxel's distance value. The distance values ​​of all voxels constituted a vascular centerline distance map with the same size as the original 3D CT angiography image of the large animal.

[0019] Extracting the centerline of the arterial vascular tree allows for precise location and analysis of the positions of various calcified lesions within the tree. This embodiment employs the fast-moving algorithm to extract the centerline, which exhibits superior computational efficiency and accuracy, making it suitable for handling complex vascular bifurcation structures. By simulating the propagation of wavefronts within the vascular lumen from one or more starting points, the fast-moving algorithm robustly generates a smooth and centered centerline.

[0020] In practice, a starting point is manually placed at the beginning of the 3D CT angiography image data sequence of large animals, for example, at the root of the ascending aorta. The fast-travel algorithm uses the starting point as the wavefront source and the voxel gray value as the velocity function. It propagates quickly in high gray value regions within the vessel lumen and slowly in low gray value regions on the vessel wall or outside. The algorithm calculates the shortest time for the wavefront to reach each voxel within the vessel, generating an arrival time map. By tracing the gradient descent path in the arrival time map, it is possible to trace back from any termination point of the arterial vascular tree to the starting point. This shortest time path constitutes the centerline of the corresponding vascular branch. For example, tracing back from the end of the iliac artery to the root of the aorta allows for the precise extraction of the centerline from the abdominal aorta to the iliac artery.

[0021] To avoid the inaccuracies and low efficiency of manually selecting seed points, in an optional embodiment, the extraction of the arterial vascular tree centerline based on image data includes: inputting the large animal 3D CT angiography image data into a keypoint detection deep learning network, identifying and outputting the 3D coordinates of the centerline start point and the centerline end point; the centerline start point corresponds to the root of the arterial vascular tree to be analyzed, and the centerline end point corresponds to the end of a branch of the arterial vascular tree; applying a vascular enhancement filter to the large animal 3D CT angiography image data to generate a vascular response map, and using the vascular response map as the velocity function of a fast traversal algorithm; using the detected centerline start point as the source point, propagating on the velocity map using the fast traversal algorithm, calculating the minimum arrival time from the source point to each voxel in the large animal 3D CT angiography image, and generating an arrival time map; starting from each detected centerline end point, backtracking along the gradient descent direction on the arrival time map until reaching the centerline start point, and the resulting path sequence constitutes the arterial vascular tree centerline.

[0022] S2, the three-dimensional CT angiography image data of the large animal and the distance map of the vessel centerline are used as multi-channel inputs and input to the pre-trained three-dimensional segmentation network to obtain the preliminary segmentation results of the arterial calcification region and the corresponding voxel-level segmentation probability map; the three-dimensional segmentation network is trained using a combined loss function including Dice loss and Focal Loss with weights adjusted according to the voxel Henle unit value.

[0023] Large animal 3D CT angiography images are used as the first channel, and the vessel centerline distance map is used as the second channel. These are stacked along the channel dimension to form a two-channel 3D image tensor. This 3D image tensor is then input into a pre-trained 3D segmentation network. The nnU-Net segmentation network can be used. The combined loss function is a linear weighted sum of Dice loss and weighted Focal Loss. The weights of the Focal Loss are set according to the Heinz unit value of the voxels. For example, voxels with a Heinz unit value higher than 130 are assigned a weight of 1.0, while voxels with a value lower than 130 are assigned a weight of 0.25, thus guiding the network to focus more on high-density regions.

[0024] In an optional embodiment, the combined loss function Defined as L=α L Dice +(1-α) L weighted_focal ,in For the loss of Des, For weighted Focal Loss, The balance coefficient is in the range [0, 1]. In the weighted Focal Loss, for any voxel, the weight is 2.0 when its Heinz unit value is greater than 130, and 0.5 otherwise.

[0025] The combined loss function integrates the advantages of Descein loss and weighted Focal loss. Descein loss effectively handles the problem of severe imbalance between positive and negative samples, which is common in calcification lesion segmentation tasks because calcified regions are very small compared to the entire scan image. Focal loss, on the other hand, focuses on directing the model to classify voxels that are difficult to classify, i.e., true calcification points with low predicted probabilities or non-calcification points with high predicted probabilities. This is achieved through a balance coefficient... This allows adjusting the contribution levels of both, for example, when When set to 0.6, it means that the model training focuses more on optimizing the overlap of segmented regions.

[0026] Arterial calcified plaques, being high-density tissue, typically have Henlein unit values ​​far exceeding those of surrounding soft tissue and blood, generally exceeding 130. Therefore, voxels with Henlein unit values ​​greater than 130 are assigned a high weight of 2.0, allowing the model to focus on learning these areas that are highly likely to be calcifications. Conversely, voxels with Henlein unit values ​​no greater than 130, such as the vascular lumen, soft tissue of the vessel wall, or image noise, are assigned a low weight of 0.5 to reduce their interference with model training. This allows the model to not only learn the contextual features of the image but also directly utilize the physical density information of the voxels, reducing misjudgments of high-density artifacts.

[0027] S3. Perform connected component analysis on the preliminary segmentation results, extract topological and morphological indicators for each candidate calcified lesion, input the topological and morphological indicators into a pre-trained Bayesian network to calculate the structure confidence score, use the structure confidence score to weight the voxel-level segmentation probability map lesion by lesion, and perform threshold processing on the weighted voxel-level segmentation probability map to obtain the segmentation results of the arterial calcification region.

[0028] Specifically, a 3D 26-connected-domain algorithm was used to analyze the preliminary segmentation results to identify each candidate calcified lesion. For each connected domain, more than ten topological and morphological indicators, such as volume, surface area, sphericity, elongation, and average centerline distance, were extracted. These indicators were then fed as node inputs into a pre-trained tree-augmented Naive Bayes network, which output a structural confidence score between 0 and 1. The voxel probability values ​​belonging to each connected domain in the voxel-level segmentation probability map corresponding to the preliminary segmentation results were multiplied by the confidence score of the corresponding connected domain to obtain a weighted voxel-level segmentation probability map. A thresholding process was then applied to the weighted voxel-level segmentation probability map to generate an arterial calcification region segmentation mask, with a threshold of 0.5.

[0029] In an optional embodiment, the topological and morphological parameters include: volume, surface area, sphericity, and maximum three-dimensional diameter.

[0030] After initially segmenting all potential calcified lesions, a set of key topological and morphological features were extracted for each connected lesion region to perform quantitative analysis and distinguish between genuine and false lesions. These features 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 summing 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 voxel surfaces at the lesion boundary; the sphericity is calculated from the volume and surface area; the maximum three-dimensional diameter refers to the maximum Euclidean distance between any two points on the lesion surface.

[0031] In an optional embodiment, the structure confidence score is used to perform lesion-by-lesion weighting on the voxel-level segmentation probability map, including: multiplying the structure confidence score of each candidate calcified lesion by the original probability value of each voxel within the candidate calcified lesion to obtain the modulated probability value.

[0032] Voxel-level segmentation probability maps only reflect the local probability that each voxel belongs to calcified tissue, but do not consider whether the entire connected region to which the voxel belongs has the typical morphology of a real calcified lesion. By introducing a structural confidence score, which comprehensively evaluates the overall shape, size, and texture of candidate lesions, a holistic judgment is provided for the authenticity of the entire lesion. For example, suppose two candidate lesions, lesion A and lesion B, are segmented. Lesion A has a regular shape and medium size, with a structural confidence score of 0.95, indicating that lesion A is very likely a real calcified lesion. The original probability value of a voxel inside lesion A is 0.8, which becomes 0.76 after modulation. Lesion B has a thin and elongated shape with blurred edges, possibly due to imaging artifacts, and a structural confidence score of only 0.2. The original probability value of a voxel inside lesion B is very high, such as 0.85, but the modulated probability value is significantly suppressed to 0.17, which can be effectively filtered out in subsequent thresholding, thus preventing misjudgment.

[0033] S4. Within the segmentation result of the arterial calcification region, calculate the high-throughput image feature set, construct the mutual information matrix between features and use the community detection algorithm to perform feature clustering, select representative features from each cluster community to form the optimal feature subset; input the optimal feature subset into the preset gradient boosting decision tree classifier to obtain the identification result of the risk level of arterial calcification status in large animals.

[0034] High-throughput image feature sets, also known as radiomics features, not only include basic morphological indicators such as lesion volume and shape, but also hundreds or thousands of refined indicators describing the internal pixel intensity distribution, spatial relationships, and complex texture patterns. Using the PyRadiomics toolkit, over one hundred high-throughput image features were extracted from large animal 3D CT angiography images within a region defined by an arterial calcification region segmentation mask. These features included first-order statistical features, gray-level co-occurrence matrix features, gray-level run-length matrix features, and gray-level region size matrix features. Normalized mutual information between all pairs of features was calculated to construct a feature mutual information matrix. Treating the feature mutual information matrix as an adjacency matrix of a weighted undirected graph, the Louvain community detection algorithm was used to cluster the high-throughput image features. For each detected community, the feature with the largest sum of mutual information values ​​with all other features within the community was selected as the representative feature of the community. All representative features together constitute the optimal feature subset. The optimal feature subset vector, composed of representative features, is input into an XGBoost gradient boosting decision tree classifier pre-trained with labeled data. Based on the input feature vector, the classifier outputs the identification result of the arterial calcification state.

[0035] In an optional embodiment, feature clustering includes: processing the mutual information matrix between high-throughput image features using the Louvain community detection algorithm, and dividing the high-throughput image feature clusters into different communities.

[0036] After extracting hundreds of high-throughput radiomics features from calcified lesions, the problems of excessive feature dimensionality and information redundancy arise. To address this, feature clustering techniques are employed to group features that are functionally or informatively similar, with mutual information used as a metric to measure the correlation between features. Specifically, the mutual information values ​​between all pairs of features are calculated, forming an N x N mutual information matrix, where N is the total number of features, for example, 500. Each element in the matrix represents the amount of information shared between the corresponding two features. This matrix can be viewed as a network graph, where features are nodes and mutual information values ​​are edge weights. The Louvain community detection algorithm is applied to this network graph, optimizing the modularity index to group features into different communities. For example, one community might cluster all features related to lesion volume and size, another might cluster features related to grayscale histogram statistics, and yet another might contain features describing the sharpness of lesion edges.

[0037] In an optional embodiment, selecting representative features from each cluster community includes: within each community, calculating the sum of mutual information values ​​between a single high-throughput image feature and all other high-throughput image features within the community, and selecting the high-throughput image feature with the largest sum as the representative feature of the community.

[0038] In an optional embodiment, the risk level of arterial calcification in large animals is identified as one of three levels: low risk, medium risk, and high risk.

[0039] Low risk corresponds to a total calcification volume of less than 10 cubic millimeters and the absence of large plaques, representing the early stage of the disease or the absence of obvious lesions. Intermediate risk corresponds to a calcification volume between 10 and 100 cubic millimeters, with scattered lesions, representing the formation of the disease but not yet reaching its most severe stage. High risk corresponds to a total calcification volume greater than 100 cubic millimeters, or the presence of confluent plaques exceeding 10 millimeters in length.

[0040] The implementation principle of the deep learning-based large animal arterial calcification CT image recognition method in this invention is as follows: This invention uses a vascular centerline distance map as network input and employs a combined loss function with weights adjusted according to Henle unit values. This enhances the segmentation network's ability to distinguish between intra-arterial calcified lesions and extra-vascular high-density tissues such as bone, improving the segmentation accuracy for small and irregularly shaped calcified lesions. Furthermore, by utilizing prior knowledge of lesion topology and morphology to perform confidence weighting and correction on the preliminary segmentation results, false positive regions caused by artifacts and other factors can be eliminated, thus ensuring high reliability of the segmentation results. Moreover, by constructing a mutual information matrix and combining it with a community detection algorithm for feature selection, the optimal feature subset with high information content and low redundancy can be efficiently selected from massive high-dimensional features. Additionally, when extracting the arterial vascular tree centerline, a keypoint detection deep learning network automates the selection of start and end points.

[0041] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A deep learning-based method for recognizing CT images of arterial calcification in large animals, characterized in that, The process includes the following steps: acquiring large animal 3D CT angiography image data; extracting the arterial vascular tree centerline based on the image data; calculating the distance from each voxel to the nearest centerline to generate a vascular centerline distance map; inputting the large animal 3D CT angiography image data and the vascular centerline distance map as multi-channel inputs into a pre-trained 3D segmentation network to obtain preliminary segmentation results of the arterial calcification region and the corresponding voxel-level segmentation probability map; the 3D segmentation network employs a Focal loss algorithm that incorporates Dice loss and weights adjusted according to the voxel Henlein unit value. The combined loss function of Loss is trained; connected component analysis is performed on the preliminary segmentation results to extract topological and morphological indicators for each candidate calcified lesion. The topological and morphological indicators are input into a pre-trained Bayesian network to calculate the structure confidence score. The structure confidence score is used to weight the voxel-level segmentation probability map lesion by lesion. Thresholding is performed on the weighted voxel-level segmentation probability map to obtain the segmentation result of the arterial calcification region; within the segmentation result of the arterial calcification region, a high-throughput image feature set is calculated. By constructing a mutual information matrix between features and using a community detection algorithm to perform feature clustering, representative features are selected from each cluster to form the optimal feature subset; the optimal feature subset is input into a preset gradient boosting decision tree classifier to obtain the identification result of the risk level of arterial calcification status in large animals.

2. The method for recognizing large animal arterial calcification CT images based on deep learning according to claim 1, characterized in that, The combined loss function is L=α L Dice +(1-α) L weighted_focal ,in For the loss of Des, For weighted Focal Loss, The balance coefficient is in the range [0, 1]. In the weighted Focal Loss, for any voxel, the weight is 2.0 when its Heinz unit value is greater than 130, and 0.5 otherwise.

3. The method for recognizing large animal arterial calcification CT images based on deep learning according to claim 1, characterized in that, The extraction of the arterial vascular tree centerline based on image data includes: The large animal 3D CT angiography image data is input into a keypoint detection deep learning network to identify and output the 3D coordinates of the centerline start point and centerline end point. The centerline start point corresponds to the root of the arterial vascular tree to be analyzed, and the centerline end point corresponds to the end of the branch of the arterial vascular tree. A vascular enhancement filter is applied to the large animal 3D CT angiography image data to generate a vascular response map, and the vascular response map is used as the velocity function of the fast traversal algorithm. Taking the detected centerline start point as the source point, the fast traversal algorithm is used to propagate on the velocity map, and the minimum arrival time from the source point to each voxel in the large animal 3D CT angiography image is calculated to generate an arrival time map. Starting from each detected centerline end point, backtracking is performed along the gradient descent direction on the arrival time map until the centerline start point is reached. The resulting path sequence constitutes the centerline of the arterial vascular tree.

4. The method for recognizing large animal arterial calcification CT images based on deep learning according to claim 1, characterized in that, The topological and morphological parameters include: volume, surface area, sphericity, and maximum three-dimensional diameter.

5. The method for recognizing large animal arterial calcification CT images based on deep learning according to claim 1, characterized in that, The step of using the structural confidence score to perform lesion-by-lesion weighting on the voxel-level segmentation probability map includes: The modulated probability value is obtained by multiplying the structural confidence score of each candidate calcified lesion by the original probability value of each voxel within the candidate calcified lesion.

6. The method for recognizing large animal arterial calcification CT images based on deep learning according to claim 1, characterized in that, The feature clustering includes: The Louvain community detection algorithm is used to process the mutual information matrix between high-throughput image features, and the high-throughput image features are clustered into different communities.

7. The method for recognizing large animal arterial calcification CT images based on deep learning according to claim 6, characterized in that, The process of selecting representative features from each cluster includes: Within each community, the sum of mutual information values ​​between a single high-throughput image feature and all other high-throughput image features within the community is calculated, and the high-throughput image feature with the largest sum is selected as the representative feature of the community.

8. The method for recognizing large animal arterial calcification CT images based on deep learning according to claim 1, characterized in that, The risk level of arterial calcification in large animals was identified as one of three levels: low risk, medium risk, and high risk.

9. The method for recognizing large animal arterial calcification CT images based on deep learning according to claim 1, characterized in that, The Bayesian network is a tree-augmented Naive Bayesian network.

10. The method for recognizing large animal arterial calcification CT images based on deep learning according to claim 1, characterized in that, The 3D segmentation network is the nnU-Net segmentation network.

Images