A coronary artery segmentation method based on hierarchical graph convolution network and double supervision mechanism
By using a hierarchical graph convolutional network and a dual supervision mechanism, combined with anisotropic diffusion filtering and branch end orientation data augmentation, the problems of broken small vessel ends and incomplete topological structure in coronary artery segmentation are solved, achieving high-precision and smooth coronary artery segmentation results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANTONG UNIV
- Filing Date
- 2026-01-20
- Publication Date
- 2026-06-05
Smart Images
Figure CN122156602A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, and in particular to a coronary artery segmentation method based on hierarchical graph convolutional networks and a dual supervision mechanism. 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. In the analysis of vascular lesions, vascular segmentation algorithms play a particularly important role. However, segmenting coronary arteries from CTA images still presents some challenges.
[0003] Previous studies have explored various methods for segmenting coronary arteries from CTA images. Methods such as region growing are relatively simple and efficient, but limitations in growth criteria hinder their accuracy and general applicability, making them unsuitable for segmenting complex vessels. Graph cut algorithms offer high accuracy but are inefficient for complex images. Level set algorithms, while computationally simple, struggle to converge effectively to target edges. Some studies have proposed using graph neural networks to predict the radius from surface points to the centerline for vessel reconstruction; however, these geometric regression-based methods often assume regular vessel cross-sections (e.g., circular or elliptical), leading to shape distortion when dealing with irregular lesions with plaques, stenosis, or bifurcation. Furthermore, existing methods often employ global random rotation or flipping during data preprocessing, neglecting the scarcity and fragility of distal branch samples, resulting in topological breaks at the distal ends of the vessels. Therefore, a segmentation method capable of sensing vascular topology and optimizing for distal breakage is urgently needed.
[0004] How to solve the above-mentioned technical problems is the challenge facing this invention. Summary of the Invention
[0005] The purpose of this invention is to provide a coronary artery segmentation method based on a hierarchical graph convolutional network and a dual-supervision mechanism. This method aims to solve the technical problems in existing technologies, such as the susceptibility of small blood vessel ends to breakage, incomplete topological structure extraction, and extreme imbalance between positive and negative samples during training. This invention can effectively improve the segmentation accuracy of small coronary arteries, particularly significantly improving the problem of topological breakage at the vessel ends. By extracting global topological features through a hierarchical graph convolutional network and a differentiable graph pooling mechanism, it achieves sub-voxel-level fine reconstruction of the blood vessel surface, ensuring the integrity and smoothness of the blood vessel structure. Simultaneously, by combining a dual-supervision mechanism and a branch-end directional data augmentation strategy, it effectively overcomes the shortcomings of high background noise interference and extreme imbalance between positive and negative samples in medical images, achieving high-precision, fully automatic end-to-end segmentation.
[0006] To achieve the aforementioned objectives, the present invention employs the following technical solution: a coronary artery segmentation method based on a hierarchical graph convolutional network and a dual supervision mechanism, comprising the following steps: Step 1: Preprocess the original 3D CTA image using anisotropic diffusion filtering technology; Step 1, image preprocessing and anisotropic diffusion filtering include the following steps: The CT values of raw CTA images typically range from -1000 HU to +3000 HU, while the CT values of coronary arteries and their contrast agents are usually distributed within a certain range. First, windowing is performed, setting the window level to 100-300 HU and the window width to 600-800 HU, mapping the region of interest to grayscale space, and removing extreme value interference from bones and lung gases.
[0007] Subsequently, to remove quantum noise from the image while preserving blood vessel edges, anisotropic diffusion filtering was employed. Its mathematical model is based on partial differential equations:
[0008] in, For image, For image gradient, For the number of iterations, This is the diffusion coefficient function. The following diffusion coefficient function is selected in this embodiment:
[0009] in This is the gradient threshold parameter. In flat regions (small gradients)... A value close to 1 indicates Gaussian smoothing; however, in the edge regions (where the gradient is large). When the value approaches 0, diffusion stops. In this embodiment, the number of iterations is set to 3, which effectively improves the sharpness of the blood vessel wall and reduces the risk of adhesion during subsequent segmentation.
[0010] Step Two: This step aims to obtain preliminary localization and a coarse mask of the coronary arteries. Due to the enormous size of 3D CTA data (e.g., 512×512×300), direct input into the network's video memory is insufficient. Therefore, a patch-based strategy is adopted.
[0011] V-Net or 3D U-Net is selected as the backbone of the coarse segmentation network. The network consists of an encoder and a decoder. The encoder extracts multi-scale semantic features through convolution and downsampling; the decoder recovers spatial resolution through upsampling and skip connections. The output layer uses the sigmoid activation function to output a probability map of each voxel belonging to a blood vessel.
[0012] The challenge of coronary artery segmentation lies in the small terminal branches. Conventional random cropping results in very few samples in the training set containing terminal vessels (class imbalance). This embodiment designs a specific augmentation strategy (branch-terminal directional data augmentation): First, based on the centerline labeled in the training set, the degree of each skeleton node is calculated. Nodes with a degree of 1 are the terminal nodes of the vascular tree. Second, local image patches of size 32×32×32 or 64×64×64 are extracted centered on these terminal nodes. Then, the tangent direction vector of the terminal blood vessel is calculated. .along The direction is Interpolation sampling is performed within the individual pixel range to generate new training sample center locations, simulating the extension of blood vessels. Finally, the tangent of the blood vessel is used. Rotate randomly by an angle around the axis. Generate multi-angle views.
[0013] Through the above operations, the proportion of samples containing highly complex and small blood vessels in the training set increased significantly, forcing the V-Net network to learn the characteristics of weak blood vessel signals and solving the problem that small blood vessels are prone to "breakage".
[0014] After training, sliding window prediction is performed on the test images, and overlapping regions are weighted and fused. Finally, a thresholding process is applied (e.g., ...). This yields a binarized, coarse vascular mask image. This forms the basis for subsequent steps.
[0015] Step 3: While the coarse mask obtained in Step 2 locates most of the blood vessels, it often contains voxel-level noise, edge spikes, or micro-fragments. To construct a high-quality graph structure, continuous and smooth centerlines must be extracted. Morphological dilation is then performed on the coarse blood vessel mask output from Step 2.
[0016] The dilation operation is described as follows: Select a 3×3×3 rectangular structuring element and align its center with every pixel in the image. For all pixels covered by the structuring element, take the maximum value. If the structuring element contains at least one foreground pixel, then the resulting pixel is set as the foreground. Repeat the above steps until all pixels in the image have been processed. This completes the image dilation operation.
[0017] The specific method for extracting the centerline using the thinning method is as follows: In 3D space, there exists a point P. Based on the Euclidean distance from this point to its neighboring points, their relationships can be categorized as 6-adjacency, 18-adjacency, and 26-adjacency. Let N(P) (j=6,18,26) represent the set of points P and their i-th neighboring points, i.e., the neighborhood. U, D, W, E, S, N, and point p represent the six directions in the 6-adjacency of Ns(P), which are the six main directions in the 3D mesh; N18(P) and six pentagrams represent N2(P). 3D thinning algorithms typically operate on binary images, which only have two values: 0 and 1. Points with a value of 1 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 a curve in the image; if N(p) contains at least one pair of relative background points, then the target point p is the endpoint of a surface 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 (by setting their value from 1 to 0) without changing the topology of the graph. They should consist only of certain types of boundary points. The endpoints of curves or surfaces are related to the topology of the graph and therefore cannot be deleted.
[0018] This invention employs a 12-direction refinement process for the 3D reconstructed blood vessels, with each iteration involving parallel refinement across all 12 directions. The order of direction selection for deletion is (UD, E, WD), (ES, UW, ND), (SW, UN, ED), and (NW, UE, SD). Based on this, the 12 directions can be divided into 4 groups, each containing 3 directions. Each subgroup includes 6 main directions. Specifically, in the first sub-iteration, certain boundary points in the U and S directions can be deleted along the US direction, and so on in subsequent iterations. The coronary artery to be refined will contract uniformly in each direction, meaning it will be eroded layer by layer from the outer layer inwards. When no more self-labeled points can be deleted in the 12 sub-iterations, the main iteration process ends, and the centerline extraction is complete.
[0019] Step 4: As a coronary artery segmentation method based on hierarchical graph convolutional networks and dual supervision mechanism provided by the present invention, in order to use graph convolutional networks to process the non-Euclidean structure of blood vessel surfaces, this step converts voxel data into graph data.
[0020] For every point on the center line Calculate its tangential vector Determine a match with An orthogonal two-dimensional cross-section. On this cross-section, with... Establish a polar coordinate system with points as poles. Use fixed angular intervals. (For example, one point every 10 degrees, for a total of 36 points) Sampling There are vertices. The position of each vertex is determined by polar coordinates. Definition, where From vertex to centerline point The Euclidean distance. Initially... The value can be obtained based on the coarse segmentation boundary from step two, or initialized as the average radius value.
[0021] The mesh depicting the surface of the blood vessel wall is represented by a graph G(V, E), where the set of nodes V represents the sampled vertices on all cross sections. The set of edges E contains two types of connections: lateral connections, where vertices at adjacent angles are connected within the same cross section (…). and Vertically connected, adjacent cross-sections are connected by vertices with the same angular index. and This forms a quadrilateral grid covering the surface of the blood vessel (which can be further subdivided into triangular grids), a structure that preserves the manifold properties of the blood vessel.
[0022] Construct node feature vectors for each graph node. Carrying a high-dimensional feature vector It integrates spatial and textural information: spatial location features are manifested in the nodes' position in the world coordinate system. The coordinates of the node and its relative coordinates with respect to the center line are used. Image texture features are represented in the preprocessed CTA image by the gray value (intensity) at the node's coordinates and the gradient magnitude of that point at different scales. To enhance the perception of boundaries, the gray values of a series of points inside and outside the node can be sampled radially as a feature profile (intensity profile).
[0023] In step five, the vascular surface mesh map constructed in step four is input into a hierarchical graph convolutional network. This network aims to predict the displacement of each node relative to the real vascular wall by learning the feature distribution of graph nodes, thereby finely adjusting the vascular surface. The hierarchical graph neural network includes a graph attention module and a differentiable graph pooling module. The graph pooling module performs hierarchical aggregation of graph nodes to extract the global topological structure features of the vascular system and outputs the corrected vascular surface node features.
[0024] Traditional GCNs (such as GCN and GraphSAGE) typically use average or max pooling when aggregating neighbor features, meaning all neighbors contribute equally to the central node. However, at coronary artery bifurcation, stenosis, or tortuosity, the local geometry is complex, and the amount of information contained in neighbors in different directions varies greatly. This embodiment introduces a self-attention mechanism. For nodes... and his neighbors Calculate the attention coefficient :
[0025]
[0026] Where || represents feature concatenation. This is the weight matrix. This is the attention vector. This mechanism enables the network to dynamically allocate aggregation weights based on the similarity of features of neighboring nodes, focusing on key node features at blood vessel bifurcation or areas with large curvature, significantly improving the segmentation accuracy of complex regions.
[0027] Coronary arteries exhibit a hierarchical tree structure. To enable the network to not only focus on local textures (such as plaques and calcifications) but also understand the global topology of the vessels (such as the hierarchical relationship between main branches and collateral branches), this embodiment embeds a DiffPool layer. Instead of directly clustering nodes, DiffPool learns an allocation matrix. , indicating the first Layer node belongs to the first The probability of a cluster.
[0028]
[0029] This matrix is used to aggregate nodes in the current layer into "supernodes" of the next layer, and the adjacency matrix is updated synchronously:
[0030]
[0031] Through multi-level pooling and unpooling operations, the network extracts a deep representation that elevates local geometric features to the overall topological features of the vascular tree. Finally, the network outputs a corrected probability or positional offset for each grid node belonging to the vascular wall. This enables subvoxel-level fine-grained reconstruction of the surface of coronary arteries.
[0032] Step 6: In order to train the above network, especially to solve the extreme class imbalance problem in blood vessel segmentation (the proportion of blood vessel voxels in the whole image is <1%), a dual loss function is constructed to train the network end-to-end and output the final result.
[0033] The dual loss function is constructed as follows:
[0034] First level of loss (Segmentation Loss): A combination of weighted cross-entropy (WCE) and Dice loss is used to address the problem of extremely small proportion of blood vessel voxels (extreme foreground-background imbalance).
[0035]
[0036]
[0037] Dice loss directly optimizes set overlap and is naturally robust to foreground-background imbalance; WCE increases foreground weights. This forces the network to focus on sparse vascular pixels.
[0038] Second loss (Vascular constraint loss): To incorporate anatomical priors, a vascular response function (Frangi Vesselness) is constructed based on the eigenvalues of the Hessian matrix. This function exhibits a high response value at tubular structures and a low response value at plate-like or spherical structures.
[0039]
[0040] As a coronary artery segmentation method based on hierarchical graph convolutional networks and dual supervision mechanism provided by the present invention, in step six, a dual loss function is constructed to train the network end-to-end and output the final result.
[0041] The dual loss function is constructed as follows:
[0042] First level of loss (Segmentation Loss): A combination of weighted cross-entropy (WCE) and Dice loss is used to address the problem of extremely small proportions of blood vessel voxels (leading to severe foreground-background imbalance). Secondary Loss (Vesselness Constraint Loss): The vascular response function is constructed based on the eigenvalues of the Hessian matrix, and the prediction results of non-tubular structures are included as a penalty term in the loss.
[0043]
[0044] in voxels Frangi vessel similarity, This constraint forces the network to predict the probability that a point is a blood vessel. This constraint forces the network output to conform to tubular geometry, which is extremely effective in suppressing non-tubular artifacts (such as veins or bone edges).
[0045] Using the enhanced training set generated in step two, the V-Net coarse segmentation network and the GCN fine segmentation network are jointly trained. The optimizer is Adam, and the initial learning rate is set to... The learning rate is adjusted using a cosine annealing strategy. During training, the Dice coefficients and Hausdorff distance (HD) on the validation set are monitored in real time, and the optimal model parameters are saved.
[0046] Ultimately, the invention takes the original CTA image as input and processes it through the above six steps in a pipeline to directly output a high-precision, topologically continuous, and surface-smooth three-dimensional coronary artery segmentation model.
[0047] Meanwhile, the present invention proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed, it implements the steps of the method described in the present invention.
[0048] Furthermore, the present invention proposes a computer-readable storage medium having a computer program stored thereon, the computer program being configured to implement the steps of the method described in the present invention when invoked by a processor.
[0049] Finally, the present invention provides a computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method described in the present invention.
[0050] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention combines the advantages of voxel-based deep learning segmentation with graph-based topology learning to design a novel 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.
[0051] 2. This invention uses a parallel network and employs two methods to segment the coronary arteries. The first segmentation branch uses VNet to perform preliminary segmentation on the enhanced block image; the second segmentation branch inputs the constructed vascular surface mesh into a Hierarchical GCN for topological refinement segmentation.
[0052] 3. This invention introduces "branch-end directional data augmentation". Unlike traditional global augmentation, this method specifically oversamples difficult samples at the ends along the tangent direction of the blood vessel, which significantly increases the network's sensitivity to small blood vessel ends.
[0053] 4. This invention introduces a differentiable graph pooling (DiffPool) mechanism, which enables the network to no longer be limited to isolated prediction of local nodes, but to learn the hierarchical topological relationship of blood vessels (such as the connection pattern of trunk and branches) through node clustering, thereby effectively repairing broken blood vessel structures.
[0054] 5. This invention constructs a dual loss function that includes vascular constraints and uses geometric prior knowledge derived from the Hessian matrix to guide the training of deep learning networks, effectively suppressing background noise and solving the problem of extreme foreground-background imbalance in coronary artery segmentation. Attached Figure Description
[0055] 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.
[0056] Figure 1 The flowchart of the coronary artery segmentation method based on hierarchical graph convolutional network and dual supervision mechanism provided by the present invention shows the complete technical path from original image input, block segmentation and centerline extraction, to the final output of segmentation results through parallel network.
[0057] Figure 2 This is a schematic diagram of 6-adjacency, 18-adjacency, and 26-adjacency connections proposed in this invention.
[0058] Figure 3 This is a schematic diagram of 12 directions in this invention.
[0059] Figure 4 This is a schematic diagram illustrating the prediction of coronary artery edge nodes and the definition of cross sections when constructing a vascular surface mesh map in the method of the present invention; Figure 4 The image shows a polar coordinate system on a two-dimensional cross-section. The method for defining blood vessel boundary nodes.
[0060] Figure 5This is a schematic diagram of the gridded blood vessel wall structure constructed in this invention; Figure 5 The image shows a three-dimensional mesh structure covering the surface of a blood vessel, formed by connecting multiple cross-sectional nodes. .
[0061] Figure 6 The images shown are examples of local vascular block images and their segmentation results involved in this invention; the left image is the extracted original block image containing the branch ends, and the right image is the corresponding segmentation mask result, demonstrating the ability of this invention to capture small branches.
[0062] Figure 7 This is a comparison chart showing the segmentation performance of the proposed method GCN+ with mainstream segmentation networks such as UNet3D, TransUNet, and NestedUNet3D, as well as the ground truth.
[0063] Figure 8 This is a comparison chart of the segmentation performance of the GCN+ method of the present invention with CS2Net, FFNet, TreeGRU segmentation networks and ground truth. Detailed Implementation
[0064] 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.
[0065] Example 1: In this example, the experiment of the present invention was implemented using Python on a Windows 10 platform. The computer platform implementing this solution has an Intel Core i5 1.6 GHz CPU and 16 GB of RAM, and an NVIDIA GeForce RTX 3090 GPU. During the experiment, the model was tested and evaluated on the ImageCAS public dataset (the ImageCAS dataset contains 1000 samples), and compared with other methods using the same dataset.
[0066] See Figures 1 to 8 This embodiment 1 proposes a coronary artery segmentation method based on hierarchical graph convolutional networks and a dual supervision mechanism; it includes the following steps: Step 1: Preprocess the original 3D CTA image using anisotropic diffusion filtering technology; Step 2: Divide the preprocessed CTA image into 3D image blocks and input them into a 3D fully convolutional neural network (such as V-Net) for voxel-level binary classification to obtain a coarse vascular mask image containing the coronary arteries. Identify the branch terminal nodes of the centerline, and perform directional data augmentation on the local image blocks containing the terminal nodes. The directional data augmentation operation includes interpolation oversampling and multi-angle rotation transformation along the tangent direction of the vessel terminal to generate an enhanced training set. Step 3: Post-process the coarse vascular mask image obtained in Step 2 to extract high-quality vascular centerlines. First, morphological dilation is performed on the coarse vascular mask to close internal voids and connect adjacent breaks, ensuring continued topological connectivity. Then, a 3D thinning algorithm is applied to iteratively erode the repaired mask until a skeleton line of monomer width, i.e., the coronary artery centerline, is obtained. Step 4: Based on the rough vascular mask obtained in Step 2 and the centerline extracted in Step 3, construct a vascular surface mesh map describing the vascular geometry. The graph nodes contain spatial location features and image texture features. Step 5: Based on the rough vascular mask obtained in Step 2 and the centerline extracted in Step 3, construct a vascular surface mesh map describing the geometric morphology of the blood vessels; the hierarchical graph neural network includes a graph attention module and a differentiable graph pooling module. The graph pooling module performs hierarchical aggregation of graph nodes to extract the global topological features of the blood vessels and outputs the corrected vascular node classification results. Step 6: Construct a composite loss function that includes segmentation loss and blood vessel morphology constraint loss, and use the enhanced training set to train the network end-to-end to output the final segmentation result.
[0067] Step 7: To verify the effectiveness and advancement of the coronary artery segmentation method based on hierarchical graph convolutional networks and a dual-supervision mechanism proposed in this invention, experiments were conducted using the publicly available large-scale coronary artery CTA dataset (ImageCAS, containing 1000 3D CTA images). The experimental environment was based on the PyTorch deep learning framework, accelerated using NVIDIA high-performance computing graphics cards, with 50 training epochs and a learning rate of 0.001. The proposed method (GCN+) was compared with mainstream segmentation networks (such as UNet3D, TransUNet, CS2Net, etc.). The Dice score, average Hausdorff distance (AHD), and Hausdorff distance (HD) were used as evaluation metrics.
[0068] The comparison results of multiple models are shown in Table 1 below:
[0069] As shown in the table above, the proposed method (GCN+) achieves a Dice score of 81.20%, significantly outperforming other comparative methods (such as UNet3D's 77.13% and TransUNet's 68.00%), demonstrating a clear advantage in overall coronary artery segmentation accuracy. Furthermore, the proposed method also exhibits excellent performance in the AHD and HD metrics, which measure boundary fit, proving the effectiveness of introducing graph convolutional networks for topology refinement.
[0070] Example 2: To further demonstrate the performance improvement of the "hierarchical graph convolutional network (GCN)" in this invention compared with the simple "block-based VNet (VNet-patch)", ablation experiment comparison results are given based on Example 1.
[0071] The model results are shown in Table 2 below:
[0072] As shown in Table 2 above, the proposed method (GCN+) improves the Dice score by approximately 5.05% and reduces the AHD by 0.6826 mm compared to the basic VNet-patch method. This fully demonstrates that by introducing a GCN network containing DiffPool pooling and attention mechanisms to refine the segmentation based on coarse segmentation, it is possible to significantly correct the topological structure breaks of blood vessels, improve the continuity and accuracy of segmentation, and reflect the technical value of the core innovation of this invention.
[0073] Furthermore, to verify the differences between different graph network structures, Table 3 shows a comparison between the present invention and the method using the GraphSAGE operator: Table 3
[0074] The results show that the hierarchical GCN structure (GCN+) adopted in this invention outperforms VNet-GraphSAGE in all metrics with only a small increase in training time (1 hour), proving the rationality and efficiency of the network structure design of this invention.
[0075] Example 3: This example proposes an electronic system, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method steps of the present invention.
[0076] Example 4: This example proposes a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of the method described in this invention, which will not be repeated here.
[0077] Example 5: This example proposes a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, they implement the steps of the method described in this invention, which will not be repeated here.
[0078] It should be noted that the processing flow of embodiments 3-4 corresponds to the specific steps of the method provided in embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the method provided in embodiment 1 of the present invention.
[0079] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.
[0080] 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 coronary artery segmentation method based on hierarchical graph convolutional networks and a dual-supervision mechanism, characterized in that, Includes the following steps: S1: Preprocess the original 3D CTA image using anisotropic diffusion filtering technology; S2: Divide the preprocessed CTA image into three-dimensional image blocks and input them into a three-dimensional fully convolutional neural network for voxel-level binary classification to obtain a coarse vascular mask image containing the coronary artery. Identify the branch terminal nodes of the centerline and perform directional data augmentation on the local image blocks containing the terminal nodes. The directional data augmentation operation includes interpolation oversampling and multi-angle rotation transformation along the tangent direction of the vascular terminal to generate an enhanced training set. S3: Post-process the rough vascular mask image obtained in S2 to extract the vascular centerline. First, perform morphological dilation operation on the rough vascular mask to close the internal cavity of the vascular and connect the adjacent fracture to ensure the continued topological connectivity. Then, apply the 3D Thinning Algorithm to iteratively erode the repaired mask until the skeleton line of the monomer width is obtained, that is, the coronary artery centerline. S4: Based on the coarse vascular mask obtained in S2 and the coronary artery centerline extracted in S3, a vascular surface mesh map describing the vascular geometry is constructed, where the graph nodes contain spatial location features and image texture features; S5: Based on the coarse vascular mask obtained in S2 and the coronary artery centerline extracted in S3, a vascular surface mesh map describing the vascular geometry is constructed; the hierarchical graph neural network includes a graph attention module and a differentiable graph pooling module. The graph pooling module performs hierarchical aggregation of graph nodes to extract the global topological features of the blood vessels and outputs the corrected vascular node classification results. S6. Construct a composite loss function that includes segmentation loss and blood vessel morphology constraint loss, and use the enhanced training set to train the network end-to-end to output the final segmentation result.
2. The coronary artery segmentation method based on hierarchical graph convolutional networks and a dual supervision mechanism according to claim 1, characterized in that, In step S1, image preprocessing and anisotropic diffusion filtering include the following steps: S11. The CT values of the original CTA images range from -1000 HU to +3000 HU. First, windowing is performed, setting the window level to 100-300 HU and the window width to 600-800 HU. The region of interest is mapped to the grayscale space to remove extreme value interference from bones and lung gases. S12. To remove quantum noise from the image while preserving blood vessel edges, anisotropic diffusion filtering is employed. Its mathematical model is based on partial differential equations: ； in, For image, For image gradient, For the number of iterations, For the diffusion coefficient function, the following diffusion coefficient function is selected: ； in The gradient threshold parameter is small in flat regions, exhibiting Gaussian smoothing; while the gradient is large in edge regions, causing diffusion to stop.
3. The coronary artery segmentation method based on hierarchical graph convolutional networks and a dual supervision mechanism according to claim 1, characterized in that, S2 includes the following steps: S21. V-Net or 3D U-Net is selected as the backbone of the coarse segmentation network. The network includes an encoder and a decoder. The encoder extracts multi-scale semantic features through convolution and downsampling. The decoder restores spatial resolution through upsampling and skip connections. The output layer uses the Sigmoid activation function to output the probability map of each voxel belonging to a blood vessel. S22. Based on the centerline labeled in the training set, calculate the degree of each skeleton node. The node with a degree of 1 is the terminal node of the vascular tree. Using this terminal node as the center, extract a local image patch with a size of 32×32×32 or 64×64×64, and calculate the tangent direction vector of the terminal blood vessel. ,along The direction is Interpolation sampling is performed within the individual pixel range to generate new training sample center positions, simulating the extension of blood vessels, and finally, the blood vessel tangent is used. Rotate randomly by an angle around the axis. Generate multi-angle views; S23. Perform sliding window prediction on the test image and perform weighted fusion of overlapping areas. Finally, after thresholding, obtain a binarized coarse blood vessel mask image.
4. The coronary artery segmentation method based on hierarchical graph convolutional networks and dual supervision mechanism according to claim 1, characterized in that, In step S3, extracting the blood vessel centerline specifically includes the following steps: S31. In three-dimensional space, firstly, based on Euclidean distance, define the 6-adjacency, 18-adjacency, and 26-adjacency relationships of points to clarify the composition of different neighborhoods. For a binary image, points with a value of 1 are target points, and points with a value of 0 are background points. By analyzing the 26-neighborhood features of target points, their type is determined: if there is only one target point in the 26-neighborhood, then that point is a curve endpoint; if there is at least one pair of opposite background points in the neighborhood, then it is a surface endpoint; all target points that have a 6-adjacency relationship with background points are boundary points. Define "simplified point" as a boundary point that is deleted without changing the topology of the graph; S32. The centerline of the reconstructed blood vessel is extracted using a 12-direction parallel thinning algorithm. The 12 thinning directions are divided into 4 groups, each containing 3 directions and covering all 6 directions. The algorithm is implemented through multiple iterations: in each main iteration, 4 sub-iterations are executed sequentially. Each sub-iteration deletes the boundary points in the corresponding direction in parallel, causing the coronary artery to contract layer by layer from the outer layer inward. When there are no new target points to delete in the 12 sub-iterations, the main iteration terminates, and the extraction of the blood vessel centerline is completed.
5. The coronary artery segmentation method based on hierarchical graph convolutional networks and dual supervision mechanism according to claim 1, characterized in that, S4 includes the following steps: S41. For each point on the central line of the coronary arteries. Calculate its tangential vector Determine a match An orthogonal two-dimensional cross section, on which, with Establish a polar coordinate system with fixed angular intervals, with the pole as the endpoint. sampling There are vertices, and the position of each vertex is determined by polar coordinates. Definition, where From vertex to centerline point The Euclidean distance, the initial The value is obtained based on the coarse segmentation boundary of S2, or initialized as the average radius value; S42. The mesh depicting the surface of the blood vessel wall is represented by G(V, E), where the set of nodes V represents the sampled vertices on all cross sections, and the set of edges E contains two types of connections: lateral connections, where vertices at adjacent angles are connected within the same cross section. and Vertically connected, adjacent cross-sections are connected by vertices with the same angle index. and This forms a quadrilateral grid covering the surface of the blood vessels; S43. Construct node feature vectors for each graph node. Carrying a high-dimensional feature vector Integrating spatial and textural information: Spatial location features are manifested in the nodes' position in the world coordinate system. The image texture features are represented by the coordinates of the node and its relative coordinates with respect to the center line. In the preprocessed CTA image, the gray value intensity at the node coordinates and the gradient magnitude of the point at different scales are also considered.
6. The coronary artery segmentation method based on hierarchical graph convolutional networks and dual supervision mechanism according to claim 1, characterized in that, S5 includes the following steps: S51. Input the vascular surface mesh map constructed in S4 into the hierarchical graph convolutional network. The hierarchical graph neural network includes a graph attention module and a differentiable graph pooling module. The graph pooling module performs hierarchical aggregation of graph nodes to extract the global topological structure features of blood vessels and outputs the corrected vascular surface node features. S52, Introduce a self-attention mechanism for nodes. and his neighbors Calculate the attention coefficient : ； ； Where || represents feature concatenation. This is the weight matrix. As the attention vector, this mechanism enables the network to dynamically allocate aggregation weights based on the similarity of features of neighboring nodes; S53. For the tree-like hierarchical structure of coronary arteries, this hierarchical topological feature is preserved in graph representation learning by embedding a DiffPool layer. This layer does not directly perform hard clustering of nodes, but learns an assignment matrix. Characterizing the first in probabilistic form Layer nodes belong to the first The correspondence between layers and clusters; ； This matrix is used to aggregate nodes in the current layer into "supernodes" of the next layer, and the adjacency matrix is updated synchronously: ； ； S54. Through multi-layer pooling and unpooling operations, the network extracts a deep representation that elevates from local geometric features to the overall topological features of the vascular tree. The network outputs the corrected probability or position offset of each grid node belonging to the vascular wall. This enables subvoxel-level fine-grained reconstruction of the surface of coronary arteries.
7. The coronary artery segmentation method based on hierarchical graph convolutional networks and dual supervision mechanism according to claim 1, characterized in that, S6 includes the following steps: S61. The dual loss function is constructed as follows: ； First level of loss : A combination of weighted cross-entropy (WCE) and Dice loss is used; ； ； Dice loss directly optimizes set overlap and is naturally robust to foreground-background imbalance; WCE increases foreground weights. This forces the network to focus on sparse vascular pixels; Second loss Based on the eigenvalues of the Hessian matrix, the Frangi Vesselness function is constructed. This function has a high response value in tubular structures and a low response value in plate-like or spherical structures. ； in voxels Frangi vessel similarity, Predict the probability that this point is a blood vessel for the network; S62. Using the enhanced training set generated in S2, jointly train the V-Net coarse segmentation network and the GCN fine segmentation network. The optimizer is Adam, and the initial learning rate is set to... In addition, the learning rate is adjusted in conjunction with the cosine annealing strategy. During the training process, the Dice coefficient and Hausdorff distance (HD) on the validation set are monitored in real time, and the optimal model parameters are saved. S63. Input the original CTA image, and after processing through steps S1 to S6 above, output a high-precision, topologically continuous, and surface-smooth three-dimensional coronary artery segmentation model.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed, it implements the steps of the method as described in any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is configured to implement the steps of the method according to any one of claims 1 to 6 when invoked by a processor.
10. A computer program product comprising a computer program / instructions, characterized in that, When executed by a processor, the computer program / instructions implement the steps of the method according to any one of claims 1 to 6.