A method for automatic 3D coronary artery segmentation based on graph attention neural network

Abstract

This invention provides an automatic 3D coronary artery segmentation method based on a graph attention neural network, belonging to the field of medical image segmentation technology. It solves the technical problems of traditional methods not fully considering the high noise and uneven grayscale in medical images. The technical solution includes the following steps: Step 1, processing the image by adjusting the window width and window level; Step 2, reducing the image size to 128×128×64; Step 3, performing dilation on the image obtained in Step 2; Step 4, establishing a graph structure based on the image obtained in Step 2; Step 5, segmenting the block image using a Vnet3D network; Step 6, fusing the coronary artery images obtained in Steps 4 and 5 to obtain the final segmented coronary artery image. The beneficial effects of this invention are: ensuring high segmentation accuracy and robustness for coronary arteries.

Description

Technical Field

[0001] This invention relates to the fields of computing and deep learning technologies, and in particular to a method for automatic segmentation of three-dimensional coronary arteries based on graph attention neural networks. Background Technology

[0002] Computed Tomography Angiography (CTA) is a non-invasive vascular imaging technique with advantages such as being non-invasive, easy to operate, and producing clear images. It has significant application value in displaying vascular lesions and vascular relationships. With the widespread use of CTA technology, vascular segmentation of CTA images has gradually become a hot topic in the field of medical image segmentation. Vascular segmentation algorithms play a particularly important role in the analysis of vascular lesions. However, segmenting coronary arteries from CTA images still presents some challenges.

[0003] Previous studies have attempted to propose methods to segment coronary arteries from CTA images. Methods such as region growing are relatively simple to compute and have high segmentation efficiency; however, due to limitations in the growth criteria, they lack good segmentation accuracy and universal segmentation rules, making them difficult to segment complex blood vessels. Graph cut algorithms can achieve high segmentation accuracy, but their efficiency is low when segmenting complex images. Level set algorithms have simple computational rules, but they struggle to converge effectively to the target edge. Solving these technical problems is the challenge of this invention. Summary of the Invention

[0004] The purpose of this invention is to provide a method for automatic segmentation of three-dimensional coronary arteries based on graph attention neural networks.

[0005] To achieve the aforementioned objectives, the present invention employs a technical solution that is fully automated, requiring no manual intervention during segmentation. The training process is end-to-end.

[0006] To achieve the above-mentioned objectives, the present invention employs the following technical solution: a three-dimensional coronary artery automatic segmentation method based on graph attention neural networks, comprising the following steps:

[0007] Step 1: Use an anisotropic diffusion filter to remove image noise; Step 1, using an anisotropic diffusion filter to remove image noise, includes the following steps:

[0008] Let I(x, 0) represent the gray level of a voxel at time t, and I0(x) be the original 3D image, div and The divergence operator and gradient operator respectively process the image as follows:

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

[0010]

[0011] The coefficient d in the above formula is a threshold used to control the diffusion rate. Specifically, it is expressed as:

[0012]

[0013] Step Two:

[0014] 1. Image reduction

[0015] Original image size: The original image size is 512×512×224. In the sagittal plane, using the image center point as the center point, the image is reduced to 128×128×64 using the resize function of the OpenCV image processing library.

[0016] 2: Coarse segmentation of Vnet3D network. Vnet3D network consists of encoder and decoder, which are composed of multiple stages.

[0017] Resolution variation: As data flows through the network, its resolution first decreases and then increases.

[0018] Feature concatenation: Similar to U-net, the output feature channel dimensions of the same level on the left are concatenated to the right to collect potentially lost fine-grained details and improve the quality of the final contour prediction.

[0019] Stage content: Each stage contains one to three convolutional layers and incorporates skip connections (pixel-level summation) to implement the residual function.

[0020] Convolutional layers: Use a 5×5×5 volume kernel (stride=1) to increase the number of channels, and use a 2×2×2 volume kernel (stride=2) to reduce the resolution.

[0021] Number of feature channels: The number of feature channels doubles at each stage of the compression path.

[0022] Activation function: PReLu.

[0023] Decoder: Employs convolutional operations and skip connections similar to the encoder, and uses 2×2 deconvolution to improve resolution.

[0024] Output: The two-channel features output are used to achieve pixel-level classification through the softmax function.

[0025] Vnet3D network segmentation process:

[0026] Input: The image reduced in size from step one.

[0027] Output: Coronary artery segmentation image.

[0028] Training parameters: The network is trained for 30 epochs, the batch size is 1, and the learning rate is set to 0.001.

[0029] 3: Enlarged segmentation results

[0030] The image segmented by the Vnet3D network is enlarged to 512×512×224 using the resize function, which is the original image size.

[0031] Step 3: Perform dilation on the image obtained in Step 2, and then use a thinning method to obtain the coronary artery centerline, specifically:

[0032] 1. Dilation operation:

[0033] Dilate the image using a 3x3x3 rectangular structuring element. Center the structuring element on each pixel and take the maximum value of all pixels covered. If the structuring element has at least one foreground pixel, set the resulting pixel as the foreground. Repeat these steps until all pixels have been processed.

[0034] 2. Refinement methods for centerline extraction:

[0035] In 3D space, based on the Euclidean distance between point P and its neighbors, adjacency is classified into 6-adjacency, 18-adjacency, and 26-adjacency. N(P) represents the neighborhood of point P, U, D, W, E, S, and N represent the six principal directions, and N18(P) and a pentagram represent 18-adjacency. Operations on a binarized image are performed, with 1 representing a target point and 0 representing a background point. If N26(P) / Pi contains exactly one target point, P is a curve endpoint; if N(P) contains at least one pair of opposing background points, P is a surface endpoint. Boundary points are at least 6-adjacent to a background point. Simplified points can be deleted without altering the graph's topology and are composed of certain types of boundary points. Curve or surface endpoints are related to the graph's topology and cannot be deleted.

[0036] 3. Three-dimensional vascular detailing:

[0037] This invention employs a 12-directional refinement method to refine the reconstructed 3D blood vessels. Each iteration involves parallel refinement along 12 directions. The deletion directions are divided into four groups, each containing three directions and six main directions. Boundary points in specific directions can be deleted during each sub-iteration. The coronary arteries contract uniformly in each direction, eroding from the outer layer inwards. Centerline extraction is complete when no more target points can be deleted after 12 sub-iterations.

[0038] Step 4: As a method for automatic 3D coronary artery segmentation based on graph attention neural network provided by the present invention, step 4 involves establishing a graph structure based on the image obtained in step 2, and segmenting the 3D coronary artery image using graph attention neural network based on the coronary artery centerline obtained in step 3. Specifically, this includes the following steps:

[0039] 1. Mesh construction: The surface mesh of the blood vessel wall is represented by graph G(V,E), where V and edge E represent the vertices and connecting edges of the mesh, respectively.

[0040] 2. Blood vessel wall and centerline: The blood vessel wall is considered as a deformable tube in 3D space. The position of each vertex is determined by the parameter rv, which is the distance from the vertex to the centerline C.

[0041] 3. Cross-section and vertex (node) positioning: Each centerline point ci corresponds to an orthogonal 2D cross-section. Vertices are distributed at equal angular intervals in the cross-section, and their positions are represented by polar coordinates (φv, rv).

[0042] 4. Mesh formation and edge connection: All cross-sectional vertices are combined to form a surface mesh G. Edges are added between adjacent vertices and between vertices with the same angle φv to form quadrilateral mesh surfaces, which are further divided into triangular surfaces.

[0043] 5. Graph Attention Network (GAT): GAT assigns weights to nodes and neighbors through a self-attention mechanism, focusing on important neighbors and ignoring irrelevant nodes.

[0044] 6. Attention Mechanism Details: The attention mechanism is a single-layer feedforward neural network that uses the LeakyReLU non-linear activation function to calculate the attention coefficients between nodes.

[0045] 7. Feature Output and Model Training: Normalized attention coefficients are used to calculate a linear combination of node features, which serves as the output feature. GAT predicts the rv value of each vertex based on image information.

[0046] Step 5: As a method for automatic 3D coronary artery segmentation based on graph attention neural networks provided by this invention, in step 5, block images of sizes 32×32×32 and 64×64×64 are obtained based on the centerline of step 3 and the original image, respectively. The Vnet3D network is then used to segment these block images. Specifically:

[0047] 1. Select a central voxel: Select a voxel from the center line as the center point.

[0048] 2. Extract block images: Using the central voxel as the center, extract cubes of sizes 32×32×32 and 64×64×64.

[0049] The block image is obtained by performing a bitwise AND operation with the original image. If the centerline voxel is located in the CTA image slice sequence...

[0050] If a cubic block image cannot be extracted from the end of a column and its center cannot be used as the center, then the slice containing the central voxel is used.

[0051] To start, expand in the opposite direction of the slice sequence to capture a complete cube-shaped block image.

[0052] 3. Input the Vnet3D network: Feed the block image as input into the Vnet3D network, with the network parameters being the same as in the previous step.

[0053] 4. Segmentation and Fusion: The Vnet3D network segments the block image and then fuses the segmentation results to generate a segmented image with the same size as the original image.

[0054] Step Six: As a method for automatic 3D coronary artery segmentation based on graph attention neural network provided by the present invention, in Step Six, the coronary artery images obtained from Step Four and Step Five are fused to obtain the final coronary artery segmentation image.

[0055] The resulting image is obtained by segmenting the Vnet3D blocks, denoted as A, and the resulting image is obtained by segmenting the graph attention, denoted as B.

[0056] 1. Define pixel values: Let A(x,y,z) and B(x,y,z) represent the pixel values ​​of images A and B at coordinates (x,y,z) in three-dimensional space, respectively.

[0057] 2. Weighted average of identical pixels: For identical pixels in images A and B, i.e., A(x,y,z) = B(x,y,z), we calculate a weighted average:

[0058] F(x,y,z)=W A ·A(x,y,z)+W B ·B(x,y,z)

[0059] Among them W A W B These are the weights of images A and B, respectively, and W... A +W B =1. Here, W is set to... A =0.5,W B =0.5.

[0060] 3. Union of distinct pixels: For distinct pixels in images A and B, i.e., A(x,y,z)≠B(x,y,z), take the union, i.e., retain the two pixel values:

[0061] F(x,y,z)={A(x,y,z),B(x,y,z)}

[0062] 4. Fusion Logical Expression: Combining the above two cases, the fused image F(x,y,z) is represented as follows:

[0063]

[0064] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0065] 1. This invention combines the feature extraction characteristics and advantages of different image segmentation methods to design a new fully automated coronary artery segmentation method. It features complete automation, requiring no manual intervention during the segmentation process. The training process is end-to-end.

[0066] 2. This invention uses a cascaded network and two methods to segment the coronary arteries. The first segmentation branch uses Vnet3D to segment the block image, and the second segmentation branch uses GAT for segmentation.

[0067] 3. In this invention, the coarse segmentation result of Vnet3D is refined to obtain the coronary artery centerline, and the coarse segmentation result is used to generate a graph structure, which is used as the input of the graph convolutional network.

[0068] 4. This invention uses a graph convolutional network (GCN) to achieve the final segmentation of the coronary arteries, obtaining the segmentation result. During the segmentation process, the GCN predicts the radius distance between the coronary artery boundary points and the center based on the center point to determine the position of the boundary points. All boundary points on the cross-section constitute the edge contour of the coronary artery. Attached Figure Description

[0069] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with Embodiments 1 and 2 of the invention to explain the invention and do not constitute a limitation thereof.

[0070] Figure 1 The overall flowchart of the automatic 3D coronary artery segmentation method based on graph attention neural network provided by the present invention is shown below.

[0071] Figure 2 This is a schematic diagram of 6-neighbor, 18-neighbor, and 26-neighbor connections proposed in step three of the automatic segmentation method for three-dimensional coronary arteries based on graph attention neural networks provided by the present invention.

[0072] Figure 3 This is a schematic diagram of the 12 directions in step three of the automatic segmentation method for three-dimensional coronary arteries based on graph attention neural networks provided by the present invention.

[0073] Figure 4This is a schematic diagram of the graph attention neural network prediction node in step four of the automatic segmentation method for three-dimensional coronary arteries based on graph attention neural networks provided by the present invention.

[0074] Figure 5 The images shown are example images of the block images obtained in step five of the automatic 3D coronary artery segmentation method based on graph attention neural network provided by the present invention. The top left and top right images are the original block images, and the bottom left and bottom right images are the block image segmentation results.

[0075] Figure 6 The images show the segmentation results of the automatic 3D coronary artery segmentation method based on graph attention neural network provided by this invention. The left image is the ground truth image provided by the dataset, and the right image is the segmentation result obtained by this invention. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and Embodiments 1 and 2. Of course, the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0077] Example 1

[0078] See Figures 1 to 6 This embodiment 1 proposes a three-dimensional automatic coronary artery segmentation method based on graph attention neural network.

[0079] In this embodiment, the experiment was implemented using Python on a Windows 10 platform. The computer platform implementing this solution has an Intel Core i5 1.6GHz CPU, 16GB of RAM, and 8GB of GPU memory. During the experiment, the model was tested and evaluated on the ImageCAS public dataset (which contains 1000 CTA image samples), and compared with other methods using the same dataset.

[0080] This embodiment includes the following steps:

[0081] Step 1: Use an anisotropic diffusion filter to remove image noise;

[0082] Step 2: Reduce the image size to 128×128×64, use the Vnet3D network to perform coarse segmentation on the reduced image, and then enlarge the segmentation result to 512×512×224;

[0083] Step 3: Perform dilation on the image obtained in Step 2, and then use a thinning method to obtain the coronary artery centerline;

[0084] Step 4: Based on the image obtained in Step 2, establish a graph structure, and based on the coronary artery centerline obtained in Step 3, segment the three-dimensional coronary artery image using a graph attention neural network;

[0085] Step 5: Based on the centerline of Step 3 and the original image, obtain block images with sizes of 32×32×32 and 64×64×64 respectively, and use the Vnet3D network to segment the block images;

[0086] Step 6: Fuse the coronary artery images obtained from Step 4 and Step 5 to obtain the final coronary artery segmentation image.

[0087] Step 7: Use FCN and Unet3D as baseline methods for experimental control. To achieve the best segmentation results for FCN and Unet3D, training is performed for 50 epochs. The learning rate is 0.001, the same as Vnet3D. The training batch size is 1, the same as Vnet3D.

[0088] The results of the three models are shown in Table 1 below:

[0089]

[0090] As can be seen from the table above, 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 achieve good coronary artery segmentation results.

[0091] Example 2

[0092] To further demonstrate the effectiveness of using parallel Vnet3D block segmentation and GAT network to segment coronary arteries in this embodiment, Example 2 is given based on Example 1 to illustrate the segmentation results in the following two different cases: A. The proposed method does not use GAT for segmentation; B. The proposed method uses Vnet3D for segmentation and simultaneously uses GAT for segmentation, and uses the fusion of Vnet block segmentation results and GAT segmentation results as the final segmentation result.

[0093] The model results are shown in Table 2 below:

[0094]

[0095] As can be seen from Table 2 above, 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 the innovation in this embodiment can improve the effect of coronary artery segmentation.

[0096] 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 principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for automatic segmentation of three-dimensional coronary arteries based on graph attention neural networks, characterized in that, Includes the following steps: Step 1: Use an anisotropic diffusion filter to remove image noise; Step 2: Reduce the image size to 128×128×64, use the Vnet3D network to perform coarse segmentation on the reduced image, and then enlarge the segmentation result to 512×512×224; Step 3: Perform dilation on the image obtained in Step 2, and then use a thinning method to obtain the coronary artery centerline; Step 4: Based on the image obtained in Step 2, establish a graph structure, and based on the coronary artery centerline obtained in Step 3, segment the three-dimensional coronary artery image using a graph attention neural network; Step 5: Based on the centerline of Step 3 and the original image, obtain block images with sizes of 32×32×32 and 64×64×64 respectively, and use the Vnet3D network to segment the block images; Step 6: Fuse the coronary artery images obtained from Step 4 and Step 5 to obtain the final coronary artery segmentation image.

2. The automatic three-dimensional coronary artery segmentation method based on graph attention neural network according to claim 1, characterized in that, Step one, using an anisotropic diffusion filter to remove image noise, includes the following steps: Let I(x, 0) represent the gray level of a voxel at time t, and I0(x) be the original 3D image, div and The divergence operator and gradient operator respectively process the image as follows: The conduction coefficient in the above formula It is a monotonically decreasing function that increases with the image gradient, defined as follows: The coefficient d in the above formula is a threshold used to control the diffusion rate.

3. The automatic three-dimensional coronary artery segmentation method based on graph attention neural network according to claim 1, characterized in that, In step two, the image is reduced to 128×128×64, and the Vnet network is used to perform coarse segmentation on the reduced image. Then, the segmentation result is enlarged to 512×512×224. Specifically, this includes the following steps: (2-1) Image preprocessing: Image resizing: Using the resize function of the OpenCV image processing library, the original image size was resized from 512×512×224 to 128×128×64; (2-2) Vnet3D Network Structure Encoder and decoder composition: The Vnet3D network consists of an encoder and a decoder, which are composed of multiple stages. The data is first reduced in resolution and then increased in resolution in the network. In the compression path, the output feature channel dimensions of the same level are concatenated to the right side to collect lost fine-grained details and improve the contour prediction quality. (2-3) Convolutional layers and skip connections Convolutional layer configuration: Each stage contains one to three convolutional layers, using 5×5×5 convolutional kernels. The size is not changed, only the number of channels is increased. A 2×2×2 convolutional kernel with stride=2 is then used to halve the resolution. The number of feature channels in each stage of the compression path is doubled, which is consistent with the design of residual networks. The activation function of the network is PReLu. Skip connections: Implementing the residual function through pixel-level summation; (2-4) Decoder Operation Decoder convolution operations: Each stage of the decoder employs convolution operations and skip connections similar to those of the encoder; Deconvolution operation: After convolution, a 2×2 deconvolution operation is used to double the resolution of each dimension; (2-5) Output and Classification Feature channel classification: The output two-channel features are classified at the pixel level using the softmax function; (2-6) Vnet Segmentation Process Input and output: The input is the scaled-down image obtained in step three, and the output is the coronary artery segmentation image; Training parameters: The network is trained for 30 epochs, the batch size is 1, and the learning rate is set to 0.001; (2-7) Image Post-processing Image size restoration: After segmentation using Vnet, the original image is resized back to 512×512×224 using the resize function, which is the original size of the original image.

4. The automatic three-dimensional coronary artery segmentation method based on graph attention neural network according to claim 1, characterized in that, In step three, the image obtained in step two is subjected to dilation, and then a thinning method is used to obtain the coronary artery centerline. Specifically, this includes the following steps: (3-1) Dilation operation: Select structural element: Select a 3×3×3 rectangular structural element; Align pixels: Align the center of the structuring element with each pixel in the image; Take the maximum value: For all pixels covered by the structuring element, take the maximum value; Foreground pixel setting: If there is at least one foreground pixel in the structuring element, the resulting pixel is set as the foreground pixel; Repeat the above steps until all pixels of the image have been processed, completing the dilation operation; (3-2) Refinement method for extracting center line A point P in three-dimensional space: Determining a point P in three-dimensional space; Adjacency relationship classification: Based on the Euclidean distance from point P to its neighboring points, the relationship is divided into 6-adjacency, 18-adjacency, and 26-adjacency; Neighborhood definition: Let N(P), j = 6, 18, 26 represent the set of adjacent points of point P and i, i.e., the neighborhood; Direction representation: U, D, W, E, S, N represent the 6 directions in the 6 adjacencies, and N18(P) and 6 pentagrams represent N2(P); Endpoint and surface endpoint determination: when N26(P) / Ni contains exactly one target; Binarization image operations: Perform operations on a binarized image, defining points with a value of 1 as target points and points with a value of 0 as background points.

5. The automatic three-dimensional coronary artery segmentation method based on graph attention neural network according to claim 1, characterized in that, In step four, a graph structure is established based on the image obtained in step two, and the three-dimensional coronary artery image is segmented using a graph attention neural network based on the coronary artery centerline obtained in step three. Specifically, this includes the following steps: (4-1) Mesh Representation and Blood Vessel Wall Model Mesh model definition: The surface mesh of the blood vessel wall is represented by a graph G(V,E), where V represents a vertex and E represents the edge connection between vertices; Assumption about the vessel wall: Assume that the vessel wall is a deformable tube in 3D Euclidean space with a known centerline C; Vertex position parameterization: The position of each vertex depends only on the parameter rv, which is the distance to the center line point; (4-2) Determining the centerline and cross-section Centerline point and cross section: For each centerline point ci, determine a 2D cross section orthogonal to the centerline direction; Cross-section vertex definition: In the cross-section plane, vertices spaced at equal angles define the cross-section of the surface mesh; (4-3) Definition of polar coordinates and mesh formation Polar coordinates definition: In a 2D plane, the position of each vertex is defined in polar coordinates (φv, rv), where φv is fixed; Polygonal cavity surface mesh formation: The combination of all vertices in all cross sections forms the polygonal cavity surface mesh G; Edge addition: Add edges between adjacent vertices in a cross section and between vertices with the same angle φv in adjacent cross sections; (4-4) Mesh Surface Definition and Subdivision Quadrilateral mesh surface definition: A quadrilateral mesh surface is defined by its four sides; Triangular facet subdivision: Each quadrilateral facet is further divided into two triangular faces; (4-5) Graph Attention Mechanism (GAT) Self-attention mechanism: GAT uses a self-attention mechanism to assign different weights to each node and its neighbors; Key focus: When aggregating neighbor information, the model should focus on more important neighbors and ignore less relevant nodes; Attention mechanism implementation: Attention mechanism a is a single-layer feedforward neural network, parameterized by weight vectors, and applies LeakyReLU nonlinearity with negative input slope α = 0.2; (4-6) Attention Coefficient Calculation and Feature Output Attention coefficient calculation: After full expansion, the coefficients calculated by the attention mechanism are represented by specific mathematical expressions, involving transpose and concatenation operations. The formula for calculating the importance of node j to node a is as follows: LeakyReLU is a non-linear activation function with a negative input slope α = 0.2, where T denotes transpose, || denotes concatenation, and the attention mechanism 'a' is a single-layer feedforward neural network parameterized by the weight vector. It is the input vector, where F is the feature dimension; Normalization and linear combination: The normalized attention coefficient is used to calculate a linear combination of its corresponding features, which serves as the final output feature of each node; (4-7) GAT Training and Prediction GAT training objective: GAT is trained to predict the rv value of each vertex based on the information of the image X encoded in the input vector xv; Predicting rv values: Through training, GAT can predict the distance rv from each vertex to the center line point, thereby assisting in the construction and analysis of the surface mesh of the blood vessel wall.

6. The automatic three-dimensional coronary artery segmentation method based on graph attention neural network according to claim 1, characterized in that, In step five, based on the centerline from step three and the original image, block images with sizes of 32×32×32 and 64×64×64 are obtained respectively. The Vnet3D network is used to segment the block images, specifically including the following steps: (5-1) Block Image Extraction Preparation Select a central voxel: Select a voxel from the center line as the center point; Define the cube sizes: Determine the sizes of the two cubes as 32×32×32 and 64×64×64 respectively; (5-2) Block Image Extraction AND operation to obtain block image: Perform an AND operation between the two cubes of the above sizes and the original image to obtain a partial image of the original image, i.e., a block image; Block image generation: Two block images of different sizes are obtained by performing a bitwise AND operation. (5-3) Vnet3D Network Input Input patch image: The extracted patch image is used as the input to the Vnet3D network; Parameter settings: The parameter settings for the Vnet3D network are the same as in step two, ensuring that the network correctly processes the input block image; (5-4) Block Image Segmentation Vnet3D network segmentation: Segmenting the input block image using the Vnet3D network; Segmentation result acquisition: After network processing, the segmentation result of each block image is obtained; (5-5) Segmentation results fusion Image fusion results: The block image results obtained by Vnet3D network segmentation are fused; Restoring the original image size: By fusing, a segmented image with the same size as the original image is obtained; (5-6) Final segmented image output Output segmented image: The final segmented image is the desired result. It is the same size as the original image, but contains detailed information after segmentation.

7. The automatic three-dimensional coronary artery segmentation method based on graph attention neural network according to claim 6, characterized in that, In step six, the coronary artery images obtained from step four and step five are fused to obtain the final coronary artery segmentation image. Specifically, the coronary artery segmentation is performed using a parallel segmentation network, and the final segmentation result image is obtained by fusing the coronary artery images obtained from step four and step five.

Images