Coronary vessel classification method and device based on improved dynamic graph convolutional network, and medium
By improving the dynamic graph convolution network combined with Transformer and DGCNN, the local topological structure and global characteristics of coronary blood vessels are extracted, and the problem of insufficient accuracy in identifying coronary blood vessel branches in the prior art is solved, and efficient vascular classification and correction are achieved , significantly improve classification accuracy and efficiency.
Patent Information
- Application Number
- CN202411010379.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-12
AI Technical Summary
The prior art is difficult to accurately identify coronary vascular branches, and traditional methods rely on artificial prior knowledge and are difficult to capture complex vascular topology, resulting in identification errors or missed identification.
The improved dynamic graph convolution network is adopted, combined with Transformer and DGCNN networks, and the network is trained through the coronary point cloud data set, local topological structures and global features are extracted, and vascular correction is performed using the minimum cost path algorithm and depth-first algorithm to ensure classification accuracy.
It significantly improves the accuracy and efficiency of coronary vascular classification, overcomes the limitations of traditional methods, can effectively deal with complex coronary vascular structures, has high practical value and broad application prospects.
Smart Images

Figure CN118941866A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and in particular to a coronary vessel classification method, device and medium based on an improved dynamic graph convolutional network. Background Art
[0002] Coronary computed tomography (CCTA) has been widely used in the diagnosis of coronary artery disease. In computer-aided diagnosis applications, the automated analysis algorithm can accurately identify the anatomical branches of the coronary arteries, which can quickly assist clinicians in assessing plaques and stenosis areas in the blood vessels and improve work efficiency.
[0003] The coronary artery is mainly composed of two blood vessels, the left coronary artery and the right coronary artery. The main branches of the left coronary artery include the left main trunk (LM), the left anterior descending artery (LAD), the left circumflex artery (LCX), and their branches, such as the branches of the left main trunk: diagonal branches (Diagonals, D), internodal branches (Septals, S), and the branches of the left circumflex artery: obtuse marginals (OM), etc. The right coronary artery includes the right coronary artery (RCA) and its branches, such as the acute marginal branches (AM), the right posterior descending artery (RPDA), the right posterior lateral branch (RPLB), etc. The anatomical structure of the coronary artery is as follows: Figure 1 shown.
[0004] Currently, automatic classification of coronary vessels faces multiple challenges that limit the accuracy and applicability of traditional methods and some deep learning techniques.
[0005] The patent "Coronary Artery Centerline Extraction Method, Device, Computer Equipment and Storage Medium" (Application Number CN202010477131.4, Publication Number CN111768411A) proposed by Wang Xu et al. describes a method for coronary artery centerline extraction. The method first segments the main branch of the coronary artery, and uses the segmentation results and the entrance and end points of the coronary artery to construct the centerline of the main branch vessel. Subsequently, the branch vessels are iteratively graded and named by calculating information such as the angle, position, and vessel length between the branch vessel and the centerline of the main branch vessel. However, due to the complex and diverse anatomical structure of coronary arteries, including individual differences such as the number of branches, angles, and morphological characteristics, traditional morphological feature algorithms may face challenges of misrecognition or missed recognition.
[0006] According to the method described in the patent "Coronary Artery Category Determination Method, Device, Electronic Device and Storage Medium" (Application No. 202210865996.7, Publication No. CN 114926700 B) by Gao Qi and Lu Yunxia, the bifurcation points of the coronary arteries are extracted to generate each segment of the blood vessel, a vascular coronary tree is constructed, and the blood vessels are classified using a classification network. However, the topological structure of the coronary arteries is complex, and traditional point cloud or two-dimensional image-based methods have difficulty capturing and processing the topological relationship of the blood vessels, which may lead to the inability to accurately represent and identify complex vascular structures. Summary of the invention
[0007] Technical purpose: In view of the defect that the existing technology cannot accurately identify coronary artery branches, the present invention discloses a coronary artery classification method, device and medium based on an improved dynamic graph convolutional network, which effectively combines the Transformer and DGCNN networks to achieve efficient extraction of local topological structure and global features of the coronary arteries, overcoming the limitation of traditional methods that rely on artificial prior knowledge.
[0008] Technical solution: In order to achieve the above technical objectives, the present invention adopts the following technical solution.
[0009] A coronary vessel classification method based on an improved dynamic graph convolutional network, the method comprising:
[0010] Construct an improved dynamic graph convolutional network and use the coronary point cloud dataset to train the improved dynamic graph convolutional network;
[0011] The coronary image to be processed is obtained, and the coronary vessel center point is obtained by image preprocessing; the coronary vessel center point is input into the improved dynamic graph convolutional network, and the vessel category of the predicted coronary vessel center point is output; and the vessel category of the coronary vessel centerline is obtained according to the predicted vessel category of the coronary vessel center point.
[0012] Preferably, after obtaining the coronary artery center points by performing image preprocessing, if the number of extracted coronary artery center points is less than a preset threshold value N, the coronary artery center points are supplemented by random repeated sampling until the total number of coronary artery center points reaches N; if the number of coronary artery center points exceeds N, the coronary artery center points are processed in batches to ensure that all coronary artery center points are effectively predicted in category.
[0013] Preferably, after the vascular category of the predicted coronary artery center point is output after improving the dynamic graph convolutional network, vascular correction is performed on the left and right coronary arteries, including: using the minimum cost path algorithm for the classified coronary artery center point, constructing a coronary vascular tree, and identifying vascular bifurcation points; and using the depth-first algorithm to construct each vascular path; and finally achieving classification correction of each vascular path.
[0014] Preferably, each segment of the vascular path is classified and corrected, including: after obtaining each segment of the vascular path, classifying all the center points on each segment of the vascular path, and uniformly assigning the category of all the coronary vessel center points on the path to the category with the highest proportion on the path.
[0015] Preferably, the process of acquiring the coronary point cloud dataset includes:
[0016] A coronary point cloud dataset is constructed based on coronary CT images as a training dataset for the improved dynamic graph convolutional network. The coronary point cloud dataset includes the coordinates of the center points of the coronary vessels and category annotations. Specifically, the following steps are included:
[0017] Acquire coronary CT images and obtain coronary vessel segmentation regions after preprocessing;
[0018] Based on the segmented area of the coronary vessels, the center point of the coronary vessels is extracted;
[0019] Each coronary artery center point is labeled with a category, and a coronary point cloud dataset including the coronary artery center point coordinates and category labels is constructed.
[0020] Preferably, the category annotation content includes: the center point of the left coronary artery and the center point of the right coronary artery; the center point of the left coronary artery and the center point of the right coronary artery are further classified and annotated, the center point of the left coronary artery includes the center point of the left main trunk, the center point of the left anterior descending branch, and the center point of the left circumflex branch; the center point of the branch of the center point of the left main trunk includes the center point of the diagonal branch and the center point of the internode branch; the center point of the branch of the center point of the left circumflex branch includes the center point of the intercircumflex branch. The center point of the right coronary artery includes the center point of the right coronary artery, and the center points of the branch of the center point of the right coronary artery include the center point of the sharp edge branch, the center point of the right posterior descending branch, and the center point of the right posterior lateral wall branch.
[0021] Preferably, the improved dynamic graph convolutional network is a dynamic graph convolutional neural network based on Transformer enhancement, with a total of 16 layers, and the structure is: input layer, TNet layer, first EdgegConv layer, first Transformer layer, second EdgeConv layer, second Transformer layer, third EdgeConv layer, third Transformer layer, first Connect layer, first MLP layer, second Connect layer, second MLP layer, third MLP layer, fourth MLP layer, fifth MLP layer, SoftMax layer; wherein, the first Transformer layer is also connected to the first Connect layer and the second Connect layer, and the second Transformer layer is also connected to the first Connect layer and the second Connect layer; wherein, the first EdgegConv layer, the first Transformer layer, the second EdgeConv layer, the second Transformer layer, the third EdgeConv layer, and the third Transformer layer are arranged in sequence, and each EdgegConv layer is connected to a Transformer layer; the first EdgegConv layer is used to receive data output by the TNet layer.
[0022] Preferably, each EdgeConv layer has the same structure, and each Transformer layer has the same structure; the calculation formula of the Transformer layer includes:
[0023] f out =connect(f Transformer ,relu(bn(convld(f Transformer -f r ×V))))
[0024] f r =V·A
[0025]
[0026] E=Q·K T
[0027] Among them, f out is the final output feature of the Transformer layer, f r is the reconstruction feature, V is the mapping matrix, f Transformeris the output feature of the previous EdgeConv layer corresponding to the Transformer layer, A is the attention matrix, Q is the query matrix, K is the key matrix, E is the intermediate matrix, sum represents the sum function; bn represents batch normalization, relu represents the activation function, connect represents the residual connection, and Convld represents one-dimensional convolution calculation; the calculation formulas of the query matrix, key matrix, and mapping matrix are:
[0028] Q = Convld(f Transformer )
[0029] K=Convld(f Transformer)
[0030] V=Convld(f Transformer )
[0031] Among them, Convld represents a one-dimensional convolution calculation.
[0032] The present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements any of the steps of the above-mentioned coronary vessel classification method based on an improved dynamic graph convolutional network.
[0033] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of any of the above-mentioned coronary vessel classification methods based on an improved dynamic graph convolutional network are implemented.
[0034] Beneficial effects:
[0035] First, the present invention applies the deep learning method based on the DGCNN network to the automatic classification of the coronary centerline, overcoming the limitation of the traditional method that relies on artificial prior knowledge;
[0036] Second, the present invention effectively combines the Transformer and DGCNN networks to achieve efficient extraction of local topological structure and global features of coronary arteries, can effectively handle the complex diversity and topological structure of coronary vessels, and combined with post-processing algorithms, significantly improves the accuracy and efficiency of coronary vessel classification, and has high practical value and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 This is a schematic diagram of the anatomical structure of the coronary vessels;
[0038] Figure 2 is a flow chart of the method of the present invention;
[0039] Figure 3It is the DGCNN-Transformer network structure diagram of the present invention;
[0040] Figure 4 Schematic diagram of the Transformer layer structure of the present invention. DETAILED DESCRIPTION
[0041] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0042] As attached Figure 2 As shown, the present invention discloses a coronary vessel classification method based on an improved dynamic graph convolutional network, comprising the following steps:
[0043] Construct an improved dynamic graph convolutional network and use the coronary point cloud dataset to train the improved dynamic graph convolutional network;
[0044] The coronary image to be processed is obtained, and the coronary vessel center point is obtained by image preprocessing; the coronary vessel center point is input into the improved dynamic graph convolutional network, and the vessel category of the predicted coronary vessel center point is output; the vessel category of the coronary vessel center line is obtained according to the predicted vessel category of the coronary vessel center point; since the coronary vessel center point is a point on the coronary vessel center line, the vessel category of the coronary vessel center point is the vessel category of the coronary vessel center line.
[0045] The image preprocessing process includes: obtaining a coronary vessel segmentation region according to the coronary artery image to be processed, and extracting a coronary vessel center point based on the coronary vessel segmentation region;
[0046] To ensure the uniformity and efficiency of the prediction process, the present invention further includes: after obtaining the coronary vessel center points by image preprocessing, if the number of extracted coronary vessel center points is less than a preset threshold value N, in the present invention, N is set to 1024, then the coronary vessel center points are supplemented by random repeated sampling until the total number of coronary vessel center points reaches N. On the contrary, if the number of coronary vessel center points exceeds N, the coronary vessel center points are processed in batches, and when the number of coronary vessel center points in each batch is less than N, the coronary vessel center points are supplemented by random repeated sampling; ensuring that all coronary vessel center points are effectively predicted in terms of category.
[0047] In addition, after the improved dynamic graph convolution network outputs the predicted vascular category of the coronary vessel center point, the present invention also includes a process of vascular correction for the left coronary artery and the right coronary artery; in the network output result, each coronary vessel center point is associated with its discrete position and category information, and the discrete position is the three-dimensional coordinate of the coronary vessel center point. In order to further improve the accuracy of the coronary centerline classification, the present invention adopts a post-processing step to refine the category information output by the improved dynamic graph convolution network.
[0048] Specifically, for the classified coronary artery center points, the minimum cost path algorithm is used to construct the coronary artery tree and identify the vascular bifurcation points; and the depth-first algorithm is used to construct each vascular path; finally, each vascular path is classified and corrected.
[0049] Below, taking the vascular correction process of the left coronary artery as an example, the steps of the vascular correction process are as follows:
[0050] Step 1: Construct the coronary vascular tree:
[0051] First, the minimum cost path algorithm is used to process the center point of the left coronary artery to construct a complete vascular tree. The vascular tree can be expressed as Tree = {P0, P1, ..., P i , ..., P n}, where P i represents the center point of the i-th coronary artery, including three-dimensional coordinate information, and n is the number of center points. The information of each node Pi includes {P parent , P child , node}, where P parent Represents the parent node of the current point, P child Represents the child node of the current point, node represents the degree of the node. The degree of the blood vessel bifurcation point is greater than 1, the degree of the blood vessel end point is 0, and the degree of the remaining blood vessel points is 1.
[0052] Step 2: Find the set of coronary artery bifurcation points:
[0053] In the constructed vascular tree, identify all nodes with degree 2 as the bifurcation point set P b = {P b1 , P b2 , ..., P bn}, where bn is the number of bifurcation points.
[0054] Step 3: Construct each vascular path
[0055] Taking each bifurcation point as the starting point, the depth-first algorithm is used to search forward and backward according to the parent node and child node information of the current point, until the next bifurcation point or end point is found. In this way, the path of each blood vessel segment is constructed. i ={p1, p2, ..., p p}, where p represents the number of points in the vessel segment. Repeat this process until all bifurcation points are searched, that is, all vessel segments are searched, and finally the set of all vessel segments Vessel = {Path1, Path2, ..., Path n}.
[0056] Step 4: Correct the category of each blood vessel segment
[0057] The categories of all the center points on each vascular path are counted, and the categories of all the coronary artery center points on the path are uniformly assigned to the category with the highest proportion on the path. In this way, the category of each vascular path is corrected to ensure the consistency of category distribution.
[0058] The right coronary artery was corrected in the same manner.
[0059] The present invention not only improves the accuracy of coronary artery centerline classification by adding a correction process, but also ensures the practicability and effectiveness of the classification results in clinical applications.
[0060] The process of acquiring the coronary point cloud data set includes:
[0061] A coronary point cloud dataset is constructed based on coronary CT images as a training dataset for the improved dynamic graph convolutional network. The coronary point cloud dataset includes the coordinates of the center points of the coronary vessels and category annotations. Specifically, the following steps are included:
[0062] Step 1: Acquire a coronary CT image, and obtain a coronary vessel segmentation area after preprocessing; including: acquiring a coronary CT image, which is a 3D image including the coronary arteries of the heart, building a coronary vessel segmentation network, cropping the acquired coronary CT image into several areas of preset size, and then inputting the obtained coronary CT image into the coronary vessel segmentation network to output the segmentation result of each area, which only contains the coronary vessel content, and splicing all the segmentation results to obtain the coronary vessel segmentation area, which is a 3D image containing the complete coronary vessel content.
[0063] In some embodiments of the present invention, the coronary vessel segmentation network uses a 3DUnet network to crop the coronary CT image into a number of 64*64*64 regions, and the cropped images are respectively input into the 3DUnet network for processing, and the segmentation results of each region are output, and all the segmentation results are spliced to obtain a complete coronary vessel segmentation region;
[0064] Step 2, extracting the center point of the coronary vessel based on the segmented coronary vessel region; comprising: obtaining a number of center points of the coronary vessel through a 3D skeleton extraction algorithm according to the segmented coronary vessel region obtained in step 1, each center point of the coronary vessel being represented by a three-dimensional point coordinate in the segmented coronary vessel region; the center point of the coronary vessel is a point located on the center line of the coronary vessel; the center point of the coronary vessel is a discrete point, and the center line of the coronary vessel is obtained by concatenating the three-dimensional point coordinates;
[0065] Step three, classify and label the center point of each coronary artery, and construct a coronary point cloud data set including the coordinates of the center points of the coronary arteries and the class labels; the class labels include: the center point of the left coronary artery and the center point of the right coronary artery; the center point of the left coronary artery and the center point of the right coronary artery are further classified and labeled, the center point of the left coronary artery includes the center point of the left main trunk, the center point of the left anterior descending branch, and the center point of the left circumflex branch, the center point of the branches of the left main trunk includes the center point of the diagonal branch and the center point of the internode branch; the center point of the branches of the left circumflex branch includes the center point of the intercircumflex branch. The center point of the right coronary artery includes the center point of the right coronary artery, and the center points of the branches of the right coronary artery include the center point of the sharp edge branch, the center point of the right posterior descending branch, and the center point of the right posterior wall branch;
[0066] As attached Figure 3 As shown in the figure, the improved dynamic graph convolutional network is a dynamic graph convolutional neural network based on Transformer enhancement, referred to as DGCNN-Transformer; specifically, it includes the following steps:
[0067] Step 4: Build a Transformer-enhanced dynamic graph convolutional neural network (DGCNN-Transformer); there are 16 layers in total, and the structure is: input layer, TNet layer, first EdgeConv layer, first Transformer layer, second EdgeConv layer, second Transformer layer, third EdgeConv layer, third Transformer layer, first Connect layer, first MLP layer, second Connect layer, second MLP layer, third MLP layer, fourth MLP layer, fifth MLP layer, SoftMax layer; among them, the first Transformer layer is also connected to the first Connect layer and the second Connect layer, and the second Transformer layer is also connected to the first Connect layer and the second Connect layer; the parameter settings of each layer include:
[0068] The input layer is used to provide input data for the TNet layer. The input data is a coronary point cloud dataset, specifically the coordinates of the center points of the coronary vessels in the coronary point cloud dataset; the size is N×3, where N represents the number of point clouds, that is, the number of center points of the coronary vessels. In the present invention, N is 1024, and 3 represents the feature dimension of the point cloud data. The feature of the point cloud data in this patent is the three-dimensional point coordinates;
[0069] The TNet layer includes three convolutional layers and three fully connected layers connected in sequence, which are used to perform affine transformation on the input data to make it translation and rotation invariant; the output dimension is N×3.
[0070] The TNet layer is a neural network layer specifically used to process point cloud data. It consists of three convolutional layers and three fully connected layers. Each convolutional layer includes convolution operations, batch normalization, and leaky ReLU activation functions. The size of the convolution kernel is set to 1×1, the step size is 1, and the number of convolution kernels is 64, 128, and 1024, respectively. Subsequently, the data processed by the three convolutional layers pass through three fully connected layers, and the number of nodes is reduced to 512, 256, and 9, respectively. The three fully connected layers map the data processed by the three convolutional layers into a 3×3 matrix through linear transformation, and finally perform matrix multiplication with the input data of the TNet layer. This process eliminates the influence of spatial transformations (such as rotation and translation) in the point cloud data on the feature extraction process and ensures the consistency of the point cloud data under different transformations, which is crucial to improving the performance of the network when processing three-dimensional data.
[0071] The first EdgeConv layer, the first Transformer layer, the second EdgeConv layer, the second Transformer layer, the third EdgeConv layer, and the third Transformer layer are arranged in sequence, and each EdgeConv layer is connected to a Transformer layer; the first EdgeConv layer is used to receive the data output by the TNet layer, and the third Transformer layer is used to output data with a dimension of N×64; each EdgeConv layer has the same structure, and each Transformer layer has the same structure;
[0072] The EdgeConv layer is used to extract local feature information of blood vessels by dynamically constructing a local topological structure. The coronary vessels are simplified into discrete coronary vessel center points. The position of the coronary vessel center points can show the morphology of the coronary vessels. Each coronary vessel center point has a coronary vessel category. The distribution of coronary vessels of different categories has certain characteristics. The EdgeConv layer is used to extract local features and find the distribution of the categories of the surrounding coronary vessel center points for each coronary vessel center point. The data processing process of the EdgeConv layer includes:
[0073] First, the K nearest neighbor algorithm (KNN) is used to determine the neighboring points of each coronary vessel center point. Specifically, with each coronary vessel center point as the center, the k nearest coronary vessel center points are found. In the present invention, k = 20. Subsequently, by calculating the position offset of the k nearest coronary vessel center points relative to the coronary vessel center point, k direction vectors are obtained, and the k direction vectors constitute the edge information in the graph structure. Finally, the k direction vectors are passed through two groups of convolutional layers to extract local feature structure information. Each convolutional layer performs convolution operation, batch normalization and Leaky ReLU activation function processing. The size of the convolution kernel is 1×1, the step size is 1, and the number of convolution kernels is 64. This design method not only improves the model's understanding of the local topological structure of the blood vessel, but also enhances the performance of the network when processing complex vascular data.
[0074] like Figure 4 As shown, the Transformer layer generates the query matrix Q, key matrix K, and mapping matrix V through three one-dimensional convolutional layers. The query matrix Q and key matrix K share weights. This process can be expressed as the following formulas (1) to (3), where f Transformer The output feature of the previous EdgeConv layer corresponding to the Transformer layer;
[0075] Q = Convld(f Transformer ) (1)
[0076] K=Convld(f Transformer ) (2)
[0077] V=Convld(f Transformer ) (3)
[0078] Among them, Convld represents a one-dimensional convolution calculation, and the number of convolution kernels of Q, K, and V are 16, 16, and 64 respectively.
[0079] Then, the intermediate matrix E is obtained by calculating the dot product of the transposed query matrix Q and the key matrix K, and the result is normalized using the sofimax function to obtain the attention matrix A, as shown in formulas (4) to (5):
[0080] E=Q·K T (4)
[0081]
[0082] Among them, sum represents the sum function;
[0083] Multiply the mapping matrix V by the attention matrix A to obtain the reconstructed feature f r, and finally through the residual connection, the output feature f Transformer And the reconstructed feature f after linear transformation r The features are added to obtain the final output feature f of the Transformer layer out , the formula is as follows:
[0084] f r =V·A (6)
[0085] f out =connect(f Transformer ,relu(bn(conv1d(f Transformer -f r ×V)))) (7)
[0086] Among them, bn represents batch normalization, relu represents activation function, and connect represents residual connection;
[0087] In the present invention, the Transformer layer is introduced into DGCNN, and the EdgeConv layer and the Transformer layer in DGCNN are arranged in sequence. This design not only retains the original feature information, but also introduces global context information, thereby improving the overall performance of DGCNN-Transformer.
[0088] The first Connect layer, the first MLP layer, the second Connect layer, the second MLP layer, and the third MLP layer are connected in sequence, and the first Connect layer simultaneously receives the output features of the first Transformer layer, the second Transformer layer, and the third Transformer layer, and the second Connect layer simultaneously receives the output features of the first Transformer layer, the second Transformer layer, and the first MLP layer; the third MLP layer is connected in series with the fourth MLP layer, the fifth MLP layer, and the SoftMax layer; the first Connect layer concatenates the output features of the first to third Transformer layers through channels. The second Connect layer concatenates the output features of the first MLP layer and the first to third Transformer layers through channels. This feature fusion method allows the network to combine feature information at different levels, allowing the network to simultaneously consider local and global features, as well as details at different scales, which is crucial for accurately identifying and analyzing coronary artery features.
[0089] The first to fifth MLP layers are used to perform deeper feature extraction on the output of the previous network layer. In the first to fifth MLP layers, each MLP layer includes a 1×1 convolution layer, batch normalization, Leaky ReLU activation function and Dropout. The number of convolution kernels in the first to fifth MLP layers is 192, 1280, 256, 128, C, respectively, where C is the number of blood vessel categories. In the present invention, C is 11, specifically LM, LAD, LCA, RCA, D, S, RI, OM, AM, RPDA, RPLB, and the Dropout parameter is set to 0.5. Dropout is a regularization technique that prevents network overfitting by randomly ignoring a part of neurons during training, thereby improving the generalization ability of the network.
[0090] As the output layer of DGCNN-Transformer, the Softmax layer is responsible for converting the features output by the previous network layer, that is, the features output by the fifth MLP layer, into a probabilistic category distribution. Specifically, the Softmax layer assigns a set of probability values to each coronary vessel center point in the coronary point cloud dataset, and the sum of these values is 1, thus ensuring that each coronary vessel center point can be clearly classified into a certain vessel category.
[0091] Step 5: Use the coronary point cloud dataset to train DGCNN-Transformer.
[0092] In the present invention, during the training phase, the stochastic gradient descent (SGD) algorithm is used as an optimizer to update the parameters of DGCNN-Transformer. The training includes 300 epochs to ensure that DGCNN-Transformer fully learns the data features. In order to improve the convergence speed of DGCNN-Transformer and ensure the stability of training, the initial learning rate is set to 0.1, and the annealing cosine function (Cosine Annealing) strategy is used to dynamically adjust the learning rate. This strategy helps DGCNN-Transformer to make more refined parameter adjustments in the later stages of training, reduce overfitting, and enhance the generalization ability of DGCNN-Transformer by periodically reducing the learning rate.
[0093] The loss function uses cross-entropy loss to measure the difference between the DGCNN-Transformer output and the true label, that is, the difference between the labeled content, to guide the training process of DGCNN-Transformer. In addition, in order to further improve the generalization ability of DGCNN-Transformer, data enhancement technology is introduced in the training process, including data translation, rotation and random perturbation operations, which can enrich the diversity of the training coronary point cloud dataset and help DGCNN-Transformer learn more robust feature representation.
[0094] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned coronary vessel classification method based on an improved dynamic graph convolutional network are implemented.
[0095] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned coronary vessel classification method based on an improved dynamic graph convolutional network are implemented.
[0096] It can be known from the description of the above implementation mode that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a general hardware platform. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product, and the computer software product can be stored in a storage medium, and the memory can be various types of memory, which can be a random access memory, a read-only memory, a flash memory, etc., such as a read-only memory (English: read-only memory, ROM) / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network communication device such as a router) to execute the methods described in each embodiment of the present application or some parts of the embodiments.
[0097] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A coronary vessel classification method based on an improved dynamic graph convolutional network, characterized in that: The method comprises: Construct an improved dynamic graph convolutional network and use the coronary point cloud dataset to train the improved dynamic graph convolutional network; The coronary image to be processed is obtained, and the coronary vessel center point is obtained by image preprocessing; the coronary vessel center point is input into the improved dynamic graph convolutional network, and the vessel category of the predicted coronary vessel center point is output; and the vessel category of the coronary vessel centerline is obtained according to the predicted vessel category of the coronary vessel center point.
2. The coronary vessel classification method based on an improved dynamic graph convolutional network according to claim 1, characterized in that: After obtaining the coronary artery center points by image preprocessing, if the number of extracted coronary artery center points is less than the preset threshold N, the coronary artery center points are supplemented by random repeated sampling until the total number of coronary artery center points reaches N; if the number of coronary artery center points exceeds N, the coronary artery center points are processed in batches to ensure that all coronary artery center points are effectively predicted.
3. The coronary vessel classification method based on an improved dynamic graph convolutional network according to claim 1, characterized in that: After the improved dynamic graph convolutional network outputs the predicted vascular category of the coronary artery center point, vascular correction is performed on the left and right coronary arteries, including: using the minimum cost path algorithm for the classified coronary artery center point to construct a coronary vascular tree and identify vascular bifurcation points; and using the depth-first algorithm to construct each vascular path; ultimately achieving classification correction of each vascular path.
4. The coronary vessel classification method based on an improved dynamic graph convolutional network according to claim 3, characterized in that: Classification and correction are performed on each segment of the vascular path, including: after obtaining each segment of the vascular path, category statistics are performed on all center points on each segment of the vascular path, and the category of all coronary vessel center points on the path is uniformly assigned to the category with the highest proportion on the path.
5. The coronary vessel classification method based on an improved dynamic graph convolutional network according to claim 1, characterized in that: The process of acquiring the coronary point cloud data set includes: A coronary point cloud dataset is constructed based on coronary CT images as a training dataset for the improved dynamic graph convolutional network. The coronary point cloud dataset includes the coordinates of the center points of the coronary vessels and category annotations. Specifically, the following steps are included: Acquire coronary CT images and obtain coronary vessel segmentation regions after preprocessing; Based on the segmented area of the coronary vessels, the center point of the coronary vessels is extracted; Each coronary artery center point is labeled with a category, and a coronary point cloud dataset including the coronary artery center point coordinates and category labels is constructed.
6. The coronary vessel classification method based on an improved dynamic graph convolutional network according to claim 5, characterized in that: The category annotation contents include: the center point of the left coronary artery and the center point of the right coronary artery; the center point of the left coronary artery and the center point of the right coronary artery are further classified and labeled, the center point of the left coronary artery includes the center point of the left main trunk, the center point of the left anterior descending branch, and the center point of the left circumflex branch, the center point of the branch of the left main trunk includes the center point of the diagonal branch and the center point of the internode branch; the center point of the branch of the left circumflex branch includes the center point of the intercircumflex branch; the center point of the right coronary artery includes the center point of the right coronary artery, and the center points of the branch of the right coronary artery include the center point of the sharp edge branch, the center point of the right posterior descending branch, and the center point of the right posterior wall branch.
7. The coronary vessel classification method based on an improved dynamic graph convolutional network according to claim 1, characterized in that: The improved dynamic graph convolutional network is a dynamic graph convolutional neural network based on Transformer enhancement, with a total of 16 layers, and the structure is: input layer, TNet layer, first EdgeConv layer, first Transformer layer, second EdgeConv layer, second Transformer layer, third EdgeConv layer, third Transformer layer, first Connect layer, first MLP layer, second Connect layer, second MLP layer, third MLP layer, fourth MLP layer, fifth MLP layer, SoftMax layer; wherein, the first Transformer layer is also connected to the first Connect layer and the second Connect layer, and the second Transformer layer is also connected to the first Connect layer and the second Connect layer; wherein, the first EdgeConv layer, the first Transformer layer, the second EdgeConv layer, the second Transformer layer, the third EdgeConv layer, and the third Transformer layer are arranged in sequence, and each EdgeConv layer is connected to a Transformer layer; the first EdgeConv layer is used to receive data output by the TNet layer.
8. The coronary vessel classification method based on an improved dynamic graph convolutional network according to claim 7, characterized in that: Each EdgeConv layer has the same structure, and each Transformer layer has the same structure; the calculation formula of the Transformer layer is include: f out =connect(f Transformer ,relu(bn(convld(f Transformer -f r ×V)))) f r =V·A E=Q·K T Among them, f out is the final output feature of the Transformer layer, f r is the reconstruction feature, V is the mapping matrix, f Transformer is the output feature of the previous EdgeConv layer corresponding to the Transformer layer, A is the attention matrix, Q is the query matrix, K is the key matrix, E is the intermediate matrix, sum represents the sum function; bn represents batch normalization, relu represents the activation function, connect represents the residual connection, and Convld represents one-dimensional convolution calculation; the calculation formulas of the query matrix, key matrix, and mapping matrix are: Q=Conv1d(f Transformer ) K=Convld(f Transformer ) V=Conv1d(f Transformer ) Among them, Convld represents a one-dimensional convolution calculation.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of a coronary vessel classification method based on an improved dynamic graph convolutional network as described in any one of claims 1-8 are implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a coronary vessel classification method based on an improved dynamic graph convolutional network as described in any one of claims 1 to 8 are implemented.
