Three-dimensional coronary artery automatic segmentation method based on graph convolutional network

Abstract

The invention provides a three-dimensional coronary artery automatic segmentation method based on a graph convolutional network, and belongs to the technical field of medical image segmentation of image segmentation. The technical problem that a traditional method does not fully consider high noise and non-uniform gray scale of a medical image is solved. According to the technical scheme, the method comprises the following steps of: 1, processing window width and window level of an image; step 2, zooming the image to the size of 128 * 128 * 64; step 3, carrying out coarse segmentation on the zoomed image by using a Vnet3D network; step 4, skeletonizing the image after coarse segmentation by using a refining method; 5, establishing a graph structure based on the image in the step 3; and step 6, based on the skeleton in the step 4 and the graph structure in the step 5, segmenting a three-dimensional coronary artery image by using a graph convolutional network. The method has the beneficial effect of ensuring high classification accuracy and robustness.

Description

Technical Field

[0001] The present invention relates to the fields of computing and deep learning technology, and in particular to a three-dimensional coronary artery automatic segmentation method based on a graph convolutional network. Background Art

[0002] Computed Tomography Angiography (CTA) is a non-invasive vascular imaging technology with the advantages of non-invasiveness, easy operation and clear imaging. It has important application value in the display of vascular lesions and vascular relationships. With the widespread application of CTA technology, vascular segmentation of CTA images has gradually become a hot issue in the field of medical image segmentation. In the analysis of vascular lesions, vascular segmentation algorithms play a particularly important role. However, there are still some challenges in segmenting coronary arteries from CTA images.

[0003] In previous studies, some scholars have tried to propose some methods to solve the problem of segmenting coronary arteries from CTA images. The proposed methods, such as the region growing method, are relatively simple to calculate and have high segmentation efficiency. However, due to the limitation of the growth criterion, the region growing method does not have good segmentation accuracy and universal segmentation criteria, and it is difficult to segment blood vessels with complex structures; the graph cut algorithm can obtain higher segmentation accuracy, but it is less efficient when segmenting complex images; the level set algorithm has simple calculation rules, but it is difficult to effectively converge to the edge of the segmentation target. How to solve the above technical problems is the subject faced by the present invention. Summary of the invention

[0004] The purpose of the present invention is to provide a three-dimensional coronary artery automatic segmentation method based on a graph convolutional network, which has the characteristics of being fully automated, does not require manual intervention during the segmentation process, and the training process is an end-to-end approach.

[0005] In order to achieve the above-mentioned invention object, the present invention adopts a technical solution: a three-dimensional coronary artery automatic segmentation method based on a graph convolutional network, comprising the following steps:

[0006] Step 1: Processing the image with window width and window position, and using an anisotropic diffusion filter to remove image noise; in the step 1, processing the image with window width and window position, and using an anisotropic diffusion filter to remove image noise includes the following steps:

[0007] First, the WW and WL set by the doctor in the original image DICOM file Tag (0028, 1050) and (0028, 1051) are read. With WL as the center, the CTA images with HU values ​​in the range of ((WL-WW) / 2, (WL+WW) / 2) are intercepted and normalized to a grayscale of 0-255. The pixels with HU values ​​less than (WL-WW) / 2 are uniformly set to 0, and the pixels with HU values ​​greater than (WL+WW) / 2 are uniformly set to 255. Then, an anisotropic diffusion filter is applied to the image to remove image noise.

[0008] Assume that I(x, 0) represents the gray level of the voxel at time t, I0(x) is the original 3D image, div and They represent the divergence operator and the gradient operator respectively, and process the image as follows:

[0009] The conduction coefficient in the above formula is is a monotonically decreasing function as the image gradient increases, defined as:

[0010]

[0011] The coefficient d in the above formula is the threshold used to control the diffusion rate. It can be specifically expressed as:

[0012]

[0013] Step 2: After the image is processed in step 1, scale the image to 128×128×64; specifically: the original image size is 512×512×224, and on the sagittal plane, with the center point of the image as the center point, use the resize() function of the OpenCV image processing library to reduce the image to 128×128×64.

[0014] Step 3: Use the Vnet3D network to roughly segment the reduced image, specifically:

[0015] The Vnet3D network consists of an encoder and a decoder, both of which consist of multiple stages. As the data passes through the network, its resolution first decreases and then increases. Similar to U-net, the output feature channel dimension of the same level on the left is spliced ​​to the right, which collects fine-grained details that may be lost in the compression path and improves the quality of the final contour prediction.

[0016] Each stage contains one to three convolutional layers, and skip connections (pixel-level summation) are added to implement the residual function.

[0017] The convolutional layer uses a volume kernel of size 5×5x5 (stride=1) (does not change the size, only increases the number of channels), and then uses a 2×2×2 volume kernel (stride=2) to achieve a similar pooling layer to reduce the resolution of each dimension to half.

[0018] The number of feature channels doubles at each stage of the compression path, and since the model is implemented as a residual network, these convolution operations double the number of feature maps while reducing the resolution.

[0019] The activation function of the network is PReLu.

[0020] At each stage of the decoder, similar convolution operations and jump connections are used as in the encoder. After the convolution, 2x2 deconvolution is used to double the resolution of each dimension.

[0021] Finally, the output two-channel features are passed through the softmax function to achieve pixel-level classification.

[0022] In the Vnet3D segmentation process, the input is the reduced image obtained in step 3, and the output is the coronary artery segmentation image. The network training iteration is 30 epochs, the batch size is 1, and the learning rate is set to 0.001.

[0023] Step 4: As a three-dimensional coronary artery automatic segmentation method based on a graph convolutional network provided by the present invention, in step 4, a refinement method is used to skeletonize the roughly segmented image, which specifically includes the following steps:

[0024] There is a point P in three-dimensional space. According to the length of the Euclidean distance from this point to its adjacent point, the relationship can be divided into 6-adjacency, 18-adjacency and 26-adjacency. N(P) (j = 6, 18, 26) represents the set of point P and i adjacent points, i.e. the neighborhood. U, D, W, E, S, N and point p represent the 6 directions in Ns(P)6-adjacency, i.e. the 6 main directions in the three-dimensional grid; N18(P) is represented by N(P) and 12 circles; N2(P) is represented by N18(P) and 6 five-pointed stars. Three-dimensional thinning algorithms usually operate on binary images, and binary images have only two values, 0 and 1. Points with a value of 1 in the image are defined as target points, and points with a value of 0 are defined as background points. If the set N26(p) / pi contains exactly one target point, then the target point P is the endpoint of the curve in the image; if N(p) contains at least one pair of relative background points, then the target point p is the surface endpoint in the image. If any target point in the image is adjacent to at least one background point 6, it is called a boundary point. Simplified points are target points that can be deleted (the value of the point is changed from 1 to 0) without changing the topological structure of the graph. They should only consist of certain types of boundary points; the endpoints of curves or surfaces are related to the topological structure of the graph and cannot be deleted.

[0025] The present invention uses 12 directions to refine the blood vessels after three-dimensional reconstruction, and each iterative step is refined in parallel by 12 directions. The order of selecting and deleting directions is (UD, E, WD), (ES, UW, ND), (SW, UN, ED), (NW, UE, SD): Based on this, the 12 directions can be divided into 4 groups, each with 3 directions. Each group contains 6 main directions, that is, in the first sub-iteration process US direction, some boundary points in the U and S directions can be deleted, and the rest of the selection processes are similar. The coronary artery that needs to be refined will shrink evenly in each direction, that is, starting from the outer layer, it will be corroded layer by layer inward. When there are no more self-marking points that can be deleted in the 12 sub-iteration processes, the main iteration process ends, that is, the centerline extraction is completed.

[0026] Step 5: As a three-dimensional coronary artery automatic segmentation method based on graph convolutional network provided by the present invention, in step 5, the segmented blood vessels are thinned, and the segmented blood vessels are used as thick blood vessels to obtain thick blood vessel skeletons. The thinning method gradually removes blood vessel edge voxels in an iterative manner until only a single voxel remains in the blood vessel area in the slice. Specifically:

[0027] The skeleton is obtained by parallel thinning method. The parallel thinning skeleton extraction algorithm repeatedly performs logical operations on the eight neighborhoods of the center pixel of each pixel to be detected, deletes non-skeleton pixels in the image, and continuously erodes and refines the extracted skeleton.

[0028] Step 6: As a three-dimensional coronary artery automatic segmentation method based on a graph convolutional network provided by the present invention, in step 6, a graph structure is established based on the image in step 3.

[0029] The mesh depicting the surface of the vessel wall is represented by a graph G(V,E) with vertices V and edges E. The vessel wall is assumed to be a deformable tube in 3D Euclidean space with a known centerline C. The position of each vertex depends only on one parameter rv. For each centerline point ci, a 2D cross section orthogonal to the centerline direction is determined. Within this cross-sectional plane, equiangularly spaced vertices define a cross section of the surface mesh. Therefore, within the 2D plane, the position of each vertex is defined in polar coordinates (φv, rv), where φv is fixed. The parameter rv is defined as the distance to the centerline point ci. The combination of all vertices in all cross sections forms a polygonal luminal surface mesh G. Edges are added between adjacent vertices in a cross section and between vertices of the same angle φv in adjacent cross sections. Four such edges define a quadrilateral mesh face, and each quadrilateral face is further divided into two triangular faces. The structure of G is fixed for a vessel or a vessel segment, and the GCN is trained to predict the rv value of each vertex based on the information of the image X encoded in the input vector xv.

[0030] Step 7: As a three-dimensional coronary artery automatic segmentation method based on a graph convolutional network provided by the present invention, in step 7, based on the center line of step 4 and the graph structure of step 5, a graph convolutional network is used to segment the three-dimensional coronary artery image.

[0031] Use a graph convolutional network to predict the value of the parameter rv given an input vector xv. GCN consists of layers that aggregate information from neighboring nodes. By connecting several such layers, information from the growing neighborhood of nodes in the graph is combined. Use the element-wise mean aggregator proposed in GraphSAGE. In the kth GCN layer, aggregate the features of vertex u in the neighborhood of vertex v, including v itself, to obtain the feature vector hk v for each node. Define h0v=xv, that is, the input of the first GCN layer is the input vector derived from image X. The aggregation rule is:

[0032]

[0033] where σ is the activation function, in this case ReLU. The matrix W contains the trainable weights that determine how features are combined between layers. Unlike GraphSAGE, neighbors are not sampled, but the full available neighbors are used. The neighborhood is defined by the edges in E, which are fixed for a given vessel segment or segment.

[0034] A network with five GCN layers is used so that the prediction for each vertex is based on vertices at most five steps away on the grid. The first layer in xv has 32 input features, and each hidden GCN layer has 64 nodes. The output layer has 1 node to predict rv. Dropout with p=0.5 is used. Given that GCN is trained using a single sample instead of mini-batches, the GCN does not contain a batch normalization layer. The GCN network contains a total of 14,567 trainable parameters. The GCN is trained using a training set consisting of coronary artery segments, each of which is represented by a graph Gi, image information Xi, and a reference value rv for all vertices v∈V in Gi. In each iteration, a segment is randomly selected and the following loss is calculated:

[0035]

[0036] Where V is the set of vertices in Gi and f(xv) is the distance to the centerline predicted by GCN. The distance value is multiplied by cube in the loss function to better ensure the correspondence between the automatically segmented vessel volume and the reference vessel volume. No additional regularization is used on the value of rv. The trained GCN can be applied to new input graphs and image data (G,X).

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. The present invention combines the feature extraction characteristics and advantages of different image segmentation methods to design a new fully automatic coronary artery segmentation method. It has the characteristics of full automation and does not require manual intervention during the segmentation process. The training process is an end-to-end method.

[0039] 2. The present invention uses a two-stage method to segment the coronary arteries. The first stage is to use Vnet3D to perform rough segmentation on the image, and the second stage is to use GCN to perform further refinement of the segmentation.

[0040] 3. In the present invention, the coarse segmentation result of Vnet3D is refined to obtain the centerline of the coronary artery, and the coarse segmentation result is used to generate a graph structure as the input of the graph convolutional network.

[0041] 4. The present invention uses a graph convolutional network to achieve the final segmentation of the coronary artery and obtain the segmentation result. During the segmentation process, GCN predicts the radius distance of the coronary artery boundary point from the center based on the center point to determine the position of the boundary point. All boundary points on the cross section constitute the edge contour of the coronary artery. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used together with Example 1 and Example 2 of the present invention to explain the present invention, but do not constitute a limitation of the present invention.

[0043] Figure 1 This is an overall flow chart of the three-dimensional coronary artery automatic segmentation method based on graph convolutional network provided by the present invention.

[0044] Figure 2 This is a schematic diagram for a detailed description of adjusting the window width and window position of an image in step 1 of the three-dimensional coronary artery automatic segmentation method based on a graph convolutional network provided by the present invention.

[0045] Figure 3 Schematic diagrams of 6-adjacency, 18-adjacency and 26-adjacency proposed in step 4 of the three-dimensional coronary artery automatic segmentation method based on graph convolutional network provided by the present invention.

[0046] Figure 4 This is a schematic diagram of 12 directions in step 4 of the three-dimensional coronary artery automatic segmentation method based on graph convolutional network provided by the present invention.

[0047] Figure 5 A schematic diagram of the graph convolutional network prediction nodes in step seven of the three-dimensional coronary artery automatic segmentation method based on graph convolutional network provided by the present invention.

[0048] Figure 6The segmentation effect diagram of the three-dimensional coronary artery automatic segmentation method based on graph convolutional network provided by the present invention; (a) is the true value diagram provided by the data set; (b) is the result diagram obtained by the segmentation of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and Example 1. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0050] Example 1

[0051] See also Figures 1 to 6 This embodiment proposes a three-dimensional coronary artery automatic segmentation method based on graph convolutional network.

[0052] In this embodiment, the experiment of this embodiment is implemented using Python language on a Windows 10 platform, and the computer platform implementing this embodiment has an Intel Core i5 1.6GHz CPU and 16GB of running memory. During the experiment, the model is tested and evaluated on the ImageCAS public dataset, which contains 1000 samples, and compared with other methods using the same dataset.

[0053] This embodiment includes the following steps:

[0054] Step 1: Process the image with different window widths and positions, and use anisotropic diffusion filter to remove image noise;

[0055] Step 2: Reduce the image to 128×128×64 size;

[0056] Step 3: Use the Vnet3D network to perform coarse segmentation on the scaled image and enlarge the resulting image to 512×512×224;

[0057] Step 4: Use the refinement method to skeletonize the roughly segmented image to obtain the centerline of the coronary artery;

[0058] Step 5: Establish a graph structure based on the image in step 3;

[0059] Step 6: Based on the skeleton of step 4 and the graph structure of step 5, a graph convolutional network is used to segment and obtain a three-dimensional coronary artery image.

[0060] Step 7: Use FCN and Unet3D as baseline methods for experimental control. Specifically, repeat steps 1 and 2, and replace Vnet3D with FCN and Unet3D in step 3 to achieve segmentation of coronary arteries. To achieve the best segmentation effect of FCN and Unet3D, train for 50 epochs. The learning rate is 0.001, which is the same as Vnet3D. The training batch size is 1, which is the same as Vnet3D.

[0061] The effects of the three models are shown in Table 1 below:

[0062]

[0063] It can be seen from Table 1 above that the method proposed in this embodiment has better results than the other two methods in terms of HD (Hausdorff Distance), AHD (Average Hausdorff Distance) and DSC (Dice Similarity Coefficient), indicating that this embodiment can achieve good coronary artery segmentation effect.

[0064] Example 2

[0065] In order to further demonstrate the influence of the extracted skeleton line on the segmentation performance of the method, Example 2 is given on the basis of Example 1 to illustrate the segmentation results in the following two different cases: A. The proposed method does not use GCN for secondary segmentation; B. The proposed method uses GCN for secondary segmentation after using Vnet3D for segmentation.

[0066] The model effect is shown in Table 2 below:

[0067]

[0068] It can be seen from Table 2 above that the method proposed in this embodiment has better results than the other two methods in terms of HD (Hausdorff Distance), AHD (Average Hausdorff Distance) and DSC (Dice Similarity Coefficient) indicators, indicating that this embodiment can improve the effect of coronary artery segmentation.

[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A three-dimensional coronary artery automatic segmentation method based on graph convolutional network, characterized in that: The following steps are involved: Step 1: Process the image with different window widths and positions, and use anisotropic diffusion filter to remove image noise; Step 2: Reduce the image to 128×128×64 size; Step 3: Use the Vnet3D network to perform coarse segmentation on the scaled image and enlarge the resulting image to 512×512×224; Step 4: Use the refinement method to skeletonize the roughly segmented image to obtain the centerline of the coronary artery; Step 5: Establish a graph structure based on the image in step 3; Step 6: Based on the skeleton of step 4 and the graph structure of step 5, a graph convolutional network is used to segment and obtain a three-dimensional coronary artery image.

2. The three-dimensional coronary artery automatic segmentation method based on graph convolutional network according to claim 1, characterized in that: In the step 1, processing the image by window width and window position, and removing image noise by using an anisotropic diffusion filter comprises the following steps: 1) Read the WW and WL set in the original image DICOM file Tag (0028, 1050) and (0028, 1051), take WL as the center, intercept the CTA image with HU value in the range of ((WL-WW) / 2, (WL+WW) / 2) and normalize it to the grayscale of 0-255, and set the pixel points with HU value less than (WL-WW) / 2 to 0, and the pixel points with HU value greater than (WL+WW) / 2 to 255; 2) Apply an anisotropic diffusion filter to the image to remove image noise; Assume that I(x, 0) represents the gray level of the voxel at time t, I0(x) is the original 3D image, div and They represent the divergence operator and the gradient operator respectively, and process the image as follows: The conduction coefficient in the above formula is It is a monotonically decreasing function as the image gradient increases, defined as follows: The coefficient d in the above formula is the threshold for controlling the diffusion rate.

3. The three-dimensional coronary artery automatic segmentation method based on graph convolutional network according to claim 1, characterized in that: In the step 2, after the processing of step 1, the image is scaled to a size of 128×128×64 to facilitate the subsequent segmentation step. Specifically, the original image size is 512×512×224. On the sagittal plane, with the center point of the image as the center point, the image is reduced to a size of 128×128×64 using the resize function of the OpenCV image processing library.

4. The method for automatic segmentation of three-dimensional coronary arteries based on graph convolutional networks according to claim 1, characterized in that: In the step 3, the scaled image is roughly segmented using the Vnet3D network, and the resulting image is enlarged to 512×512×224, including the following steps: (1) Feature detail preservation: Designed to preserve the fine-grained features lost in the process and improve the quality of the final contour prediction; (2) Encoder architecture: Each stage contains 1 to 3 convolutional layers, with skip connections added to implement the residual function. The convolutional layer uses a 5×5×5 volume kernel with stride=1 and a 2×2×2 volume kernel for downsampling to reduce the resolution of each dimension. (3) Feature channel changes: At each stage of the compression path, the number of feature channels doubles. Due to the residual network, the convolution operation increases the number of feature maps while reducing the resolution. (4) Activation function: The network uses PReLu as the activation function; (5)Decoder architecture: The decoder uses similar convolution operations and jump connections as the encoder at each stage. After the convolution, a 2×2 deconvolution is used to double the resolution of each dimension. (6) Output processing: Use the softmax function to process the output two-channel features to achieve pixel-level classification; (7) Training process: The input is the reduced image after step 3, and the output is the coronary artery segmentation image. The network training iterates for 30 epochs, the batch size is 1, and the learning rate is set to 0.

001.

5. The three-dimensional coronary artery automatic segmentation method based on graph convolutional network according to claim 1, characterized in that: In the step 4, skeletonizing the roughly segmented image using a refinement method to obtain the coronary artery centerline specifically includes the following steps: ① Adjacency definition: According to the Euclidean distance from point P to its adjacent points in three-dimensional space, the relationship is divided into 6-adjacency, 18-adjacency and 26-adjacency. N(P) represents the set of points P and i adjacent points j = 6, 18, 26; ②Direction signs: In 6-adjacency, point P is adjacent to six directions (U, D, W, E, S, N). Use N(P) and 12 circles to represent N18(P), and use N18(P) and 6 five-pointed stars to represent N26(P); ③ Binarization image processing: A binary image has only two values: 0 and 1, where the point with a value of 1 is defined as the target point and the point with a value of 0 is defined as the background point. ④Curve endpoint recognition: If the set N26(P) contains a target point, then the target point P is considered as a curve endpoint. If N(P) contains at least one pair of relative background points, then P is a surface endpoint. If any target point in the image is 6-adjacent to at least one background point, it is called a boundary point. ⑤Simplified point definition: Simplified points are target points that can be deleted without changing the topological structure of the graph. They should only consist of boundary points. The endpoints of curves or surfaces are related to the topological structure of the graph and cannot be deleted. ⑥Detailed treatment plan: The blood vessels after 3D reconstruction were refined by parallel refinement in 12 directions. The order of deleting directions was (UD, E, WD), (ES, UW, ND), (SW, UN, ED), and (NW, UE, SD). The 12 directions were divided into 4 groups, each containing 3 directions and 6 main directions. (7) Refinement process: In each sub-iteration, some boundary points in the U and S directions are deleted from the US direction, and the coronary artery shrinks uniformly in each direction, corroding inward layer by layer; (8) End of the main iteration: When there are no more target points that can be deleted in the 12 sub-iteration processes, the main iteration process ends and the centerline extraction is completed.

6. The three-dimensional coronary artery automatic segmentation method based on graph convolutional network according to claim 1, characterized in that: In the step 5, a graph structure is established for the image based on the step 3; the steps include: S51. Define the grid graph G: G is used to depict the blood vessel wall and consists of vertices V and edges E; S52, Vessel Wall and Centerline: The vessel wall is considered as a 3D deformable tube with a known centerline C and vertex positions determined by the parameter rv; S53, cross section and vertex: For each centerline point ci, an orthogonal 2D cross section is determined. Within the cross section, equiangularly spaced vertices define a grid whose position is given by polar coordinates (φv, rv), where φv is fixed and rv is the distance to ci; S54. Form a mesh and add edges: Combine all cross-section vertices to form a polygonal inner cavity surface mesh G, and add edges between adjacent vertices of the cross section and between vertices of the same angle φv in adjacent cross sections; S55, mesh surface and GCN training: The quadrilateral mesh surface is divided into two triangular surfaces. GCN predicts the rv value of each vertex based on image information, and the mesh structure is fixed to the blood vessel segment.

7. The method for automatic segmentation of three-dimensional coronary arteries based on graph convolutional networks according to claim 6, characterized in that: In step 6, based on the skeleton of step 4 and the graph structure of step 5, a graph convolutional network is used to segment the three-dimensional coronary artery image, which includes the following steps: S61, GCN network overview: Use graph convolutional network GCN to predict the parameter rv value of each node in the graph. GCN consists of multiple layers, each layer aggregates information from neighboring nodes. By connecting several such layers, information from the growing node neighborhood in the graph is combined; Working principle of S62 and GCN layer: In the kth GCN layer, the features of vertex v and vertex u in its neighborhood, including v itself, are aggregated to form a feature vector hkv. It is defined that h0v = xv, that is, the input of the first layer of GCN is the input vector derived from the image X, and the element-wise mean aggregator in GraphSAGE is used for aggregation. The aggregation rule contains the activation function σ and the trainable weight matrix W. The aggregation rule is: S63, GCN layer configuration: No sampling of neighbors is performed, and the full available neighbors are used. The neighborhood is defined by the edge set E, which is fixed for a given vessel segment. Five GCN layers are used so that predictions are based on vertices at most five steps away on the grid. The first layer has 32 input features, each hidden layer has 64 nodes, and the output layer has 1 node predicting rv; S64. Network training and parameters: Using Dropout with p=0.5, the GCN network contains a total of 14,567 trainable parameters due to single sample training and no batch normalization layer; S65, training set and loss function: GCN is trained using a training set consisting of coronary artery segments. Each coronary artery segment includes a graph Gi, image information Xi, and a reference value rv. In each iteration, a segment is randomly selected and the loss is calculated. In each iteration, a segment is randomly selected and the following loss is calculated: Where V is the vertex set in Gi, f(xv) is the distance to the centerline predicted by GCN, the loss function is based on the centerline distance predicted by GCN, the distance value is multiplied by cube to optimize the vessel volume correspondence, and no additional regularization is used for the rv value; S66, GCN applications: The trained GCN is applied to the new input map and image data (G, X). The centerline and grid structure data are input into the GCN to iteratively segment each slice. Finally, the coronary artery contours obtained from each segmentation constitute a complete coronary artery.

Images