Blood vessel continuous segmentation method based on graph network

Abstract

The invention discloses a blood vessel continuous segmentation method based on a graph network, and relates to the technical field of blood vessel continuous segmentation, multi-scale texture features based on coronary artery influence and topological structure features of blood vessels are fused, and correlation among different features is enhanced through an attention mechanism; segmenting the fused multi-scale texture features influenced by the coronary artery and topological structure features of the blood vessel, extracting multi-scale features, and reducing the resolution; gradually recovering the resolution by using a decoder to obtain a segmentation result, and optimizing the network weight by using an error between the segmentation result and a real label; and applying the trained segmentation network to test data to obtain a three-dimensional segmentation result of the coronary artery, and evaluating the accuracy of the segmentation result by comparing the difference between a predicted value and a true value to obtain connection constraint loss. The problem that a three-dimensional blood vessel structure is difficult to extract by a general medical segmentation model is solved, so that a blood vessel segmentation result is more continuous and more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of vascular continuous segmentation, and specifically to a method for vascular continuous segmentation based on a graph network. Background Art

[0002] Vascular segmentation, especially coronary artery segmentation, is an important task in medical imaging. Accurate vascular segmentation plays a crucial role in the diagnosis, treatment, and surgical planning of cardiovascular diseases. With the continuous progress of medical imaging technology, three-dimensional image data (such as CT or MRI) has become an important diagnostic tool. However, vascular segmentation still faces many challenges, especially due to problems such as complex vascular morphology, high noise in image data, and large differences in vascular structures among different patients.

[0003] Most traditional vascular segmentation methods are based on two-dimensional images or rely on simple threshold or region growing methods. These methods often have difficulty in dealing with complex structures, different sizes and shapes of blood vessels, and may be sensitive to noise and irregular shapes. Some methods only rely on single-scale image features or simple morphological feature extraction, fail to fully utilize the multi-scale texture features and topological structure features of coronary arteries, and existing vascular segmentation methods fail to effectively fuse different features and lack intelligent processing of the relationships between features. At the same time, traditional vascular segmentation networks have deficiencies in restoring details, resulting in the accuracy of segmentation results being affected.

[0004] Therefore, in view of the above problems, there is an urgent need for a method for vascular continuous segmentation based on a graph network. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a method for vascular continuous segmentation based on a graph network, which solves the problem that general medical segmentation models are difficult to extract three-dimensional vascular structures, and makes the vascular segmentation results more continuous and accurate.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for continuous segmentation of blood vessels based on a graph network, comprising the following steps: obtaining three-dimensional coronary artery image data and constructing a coronary artery image data set; extracting multi-scale texture features of the coronary artery image based on the coronary artery image data set, and extracting topological structure features of the blood vessels based on a coronary artery topological feature extraction module; fusing the multi-scale texture features of the coronary artery image and the topological structure features of the blood vessels, and enhancing the correlation between different features through an attention mechanism; using an encoder to segment the fused multi-scale texture features and topological structure features of the coronary artery image, thereby extracting multi-scale features and reducing the resolution; using a decoder to gradually restore the resolution to obtain a segmentation result, and then calculating the error between the segmentation result and the true label, and optimizing the network weights based on the error; applying the trained segmentation network to test data to obtain a three-dimensional segmentation result of the coronary artery, and evaluating the accuracy of the segmentation result by comparing the gap between the predicted value and the true value, resulting in the loss of connection constraints.

[0007] Further, the specific analysis of obtaining three-dimensional coronary artery image data and constructing a coronary artery image data set is as follows: collecting multiple three-dimensional coronary artery angiography images for different individuals using a computed tomography device, and storing the multiple three-dimensional coronary artery angiography images in a standard medical image format to obtain three-dimensional coronary artery image data; preprocessing the three-dimensional coronary artery image data according to the resolution, size, and image quality of the three-dimensional coronary artery image data, and the preprocessing includes noise removal and image normalization operations; constructing coronary artery image data sets for training and testing respectively, and the coronary artery image data set includes three-dimensional coronary artery image data of multiple different individuals, covering various anatomical variations of the coronary artery.

[0008] Furthermore, the specific analysis of extracting multi-scale texture features of coronary artery images based on the coronary artery image dataset and extracting the topological structure features of blood vessels based on the coronary artery topological feature extraction module is as follows: The coronary artery image dataset is input into the network, and then through the windowed self-attention mechanism and convolutional operations in Swin Transformer, feature information at different levels is extracted layer by layer. At different levels of the encoder, by fusing low-level and high-level features, richer texture information is extracted to obtain multi-scale texture features of coronary artery images. The multi-scale texture features of coronary artery images include local texture features and global texture features; a graph structure is constructed based on the blood vessel network of the coronary artery, with each blood vessel branch point as a node in the graph and the connection mode of the blood vessels as the edges in the graph; according to the spatial distribution of the coronary artery blood vessels, the blood vessel branch nodes are automatically located, and the edges of the graph are formed through the connection relationships; a graph convolutional network is used to perform feature propagation on each node and edge in the graph to extract the topological structure features of the blood vessels. The topological structure features of the blood vessels include the connectivity of the blood vessels, the depth of the branches, the topological type of the blood vessels, the blood flow direction of each branch, the density of the blood vessel network, and the structural stability.

[0009] Furthermore, the specific analysis of fusing the multi-scale texture features of coronary artery images and the topological structure features of blood vessels and enhancing the correlation between different features through the attention mechanism is as follows: The attention mechanism is used to enhance the representation of important features and suppress unimportant features, and then the correlation between the multi-scale texture features of coronary artery images and the topological structure features of blood vessels is learned, and appropriate weights are assigned to each feature; the feature fusion module takes the multi-scale texture features of coronary artery images and the topological structure features of blood vessels as inputs, and combines the weights of each feature assigned by the attention mechanism to form a comprehensive feature; the fused comprehensive feature vector is normalized and dimension-reduced, and then transmitted to the subsequent learning model.

[0010] Furthermore, the specific analysis of using the encoder to segment the multi-scale texture features of coronary artery images and the topological structure features of blood vessels after fusion, and then extracting multi-scale features and reducing the resolution is as follows: The fused comprehensive feature is extracted, and then based on the segmentation requirements, the fused comprehensive feature is segmented into multiple small blocks, and each small block contains multi-scale texture features and topological features; a convolutional neural network is used to perform convolutional operations to map the features in each small block to a higher-level feature space; the spatial resolution of the features is reduced through pooling operations while retaining important information. At the same time, the encoder extracts the hierarchical features in the comprehensive feature and reduces the resolution through step-by-step convolutional and pooling operations; the features at different scales are fused to form a multi-scale comprehensive feature map, and then a dimension reduction technique is used to reduce the complexity of the high-dimensional features and map the high-dimensional features to a low-dimensional space to form a low-resolution fused multi-scale feature.

[0011] Further, the decoder is used to gradually restore the resolution to obtain the segmentation result. Then, the error is calculated using the segmentation result and the ground truth label, and the network weights are optimized based on the error. The specific analysis is as follows: Receive the ground truth label corresponding to the fused multi-scale features of low resolution, where the ground truth label is the actual segmentation result of the image and is used to train and optimize the network. Gradually increase the spatial dimension of the fused multi-scale features of low resolution through upsampling or deconvolution operations until the fused multi-scale features of low resolution are restored to the image size of the original resolution, and obtain the intermediate results output at each step during the process of gradually increasing the spatial dimension. The intermediate results are specifically the gradually restored segmentation prediction maps. Use the loss function in real time to compare each intermediate result output at each step with the ground truth label to obtain the loss value between each intermediate result output at each step and the ground truth label. The loss value reflects the difference between the network output and the true segmentation label. Transmit the loss value back to each layer of the network through the backpropagation algorithm, calculate the gradients of each layer, and then update the weight parameters of the network using the optimization algorithm based on the calculated gradients. Specifically, update the weight parameters of the network based on the minimization of the loss value.

[0012] Further, apply the trained segmentation network to the test data to obtain the three-dimensional segmentation result of the coronary artery, and evaluate the accuracy of the segmentation result by comparing the gap between the predicted value and the true value. The specific analysis of the connection constraint loss is as follows: Load the updated network weight parameters obtained during the process of the decoder gradually restoring the resolution. Obtain the test data in the coronary artery image dataset, where the test data includes images and corresponding ground truth labels. Standardize the test images and adjust them to the input size used during training. Input the adjusted test images into the trained segmentation network for forward propagation to obtain the segmentation prediction results of the coronary artery. Recombine the segmentation prediction results of each slice to obtain the complete three-dimensional segmentation result, and evaluate the segmentation performance of the model based on the difference between the predicted segmentation result and the ground truth label. Detect whether there are discontinuous regions in the three-dimensional segmentation result based on the connected regions in the three-dimensional segmentation result, that is, whether there is a connection loss situation. If a connection loss situation is detected, use morphological dilation operations or graph-based optimization methods to restore the lost regions.

[0013] The present invention has the following beneficial effects: By extracting multi-scale texture features and topological structure features and fusing them, it is possible to capture the detailed information of coronary artery vessels more comprehensively. The multi-scale features help improve the segmentation ability for vessels of different sizes, while the topological structure features help understand the connection and structural relationship of vessels, thereby improving the segmentation accuracy; By introducing an attention mechanism to enhance the correlation between different features, it helps to automatically focus on key features during the vessel segmentation process, suppress irrelevant information, reduce the situation of mis-segmentation, and improve the performance of the model, especially in the segmentation task in complex scenarios; The encoder is responsible for dimensionality reduction and feature extraction, and the decoder gradually restores the resolution. This architecture has been proven effective in medical image segmentation. In this way, it is possible to recover from rough features to high-precision segmentation results, especially in the processing of 3D vessel data, and provide more accurate segmentation; Calculating the error using the segmentation result and the ground truth label and optimizing the network weights based on the error helps to improve the accuracy and robustness of the model, enabling the model to continuously improve, reduce bias, and provide segmentation results that better meet the actual clinical needs; Processing the three-dimensional coronary artery image data enables vessel segmentation not only to be limited to two-dimensional slices, but also to provide continuous and accurate vessel segmentation results in three-dimensional space, which is of great significance in the diagnosis and surgical planning of cardiovascular diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flowchart of a method for continuous vessel segmentation based on a graph network according to the present invention.

[0015] Figure 2 It is a flowchart of the core segmentation steps of a method for continuous vessel segmentation based on a graph network according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] An embodiment of the present application adopts a method for continuous vessel segmentation based on a graph network, which can accurately extract the tree-like structure in the coronary artery image and achieve high continuity and high accuracy of the segmentation result.

[0017] The general idea of the embodiment of the present application is: Through multi-stage feature extraction and fusion, combined with the network architecture of deep learning (especially the encoder-decoder structure) and the attention mechanism, accurately extract the multi-scale texture features and topological structure features of the vessels from the three-dimensional coronary artery image data, and then perform accurate three-dimensional continuous segmentation on the coronary arteries.

[0018] Please refer to Figure 1, an embodiment of the present invention provides a technical solution: a method for continuous segmentation of blood vessels based on a graph network, including the following steps: obtaining three-dimensional coronary artery image data and constructing a coronary artery image dataset; extracting multi-scale texture features of the coronary artery based on the coronary artery image dataset, and extracting topological structure features of the blood vessels based on a coronary artery topological feature extraction module; fusing the multi-scale texture features of the coronary artery and the topological structure features of the blood vessels, and enhancing the correlation between different features through an attention mechanism; using an encoder to segment the fused multi-scale texture features and topological structure features of the coronary artery, and then extracting multi-scale features and reducing the resolution; using a decoder to gradually restore the resolution to obtain a segmentation result, and then calculating the error between the segmentation result and the ground truth label, and optimizing the network weights based on the error; applying the trained segmentation network to test data to obtain a three-dimensional segmentation result of the coronary artery, and evaluating the accuracy of the segmentation result by comparing the gap between the predicted value and the ground truth value, resulting in the loss of connection constraints.

[0019] Specifically, please refer to Figure 2 , the core segmentation steps of the embodiment of the present invention are specifically analyzed as follows:

[0020] S1, obtaining a rough segmentation result using SWIN-UNetr;

[0021] S2, a blood vessel mapping module, which includes node feature extraction and graph connectivity constraint;

[0022] S3, predicting the connectivity of the third-step image, and combining the losses of the previous two steps to obtain a fine segmentation result of the patch;

[0023] S4, stitching different patches to obtain an overall segmentation result of a single CTA.

[0024] In this implementation, in S1, according to SWIN-UNetr, the rough segmentation result and loss are obtained, and the calculation method is as follows:

[0025] where p i and g i respectively represent the predicted probability and the ground truth value of voxel i.

[0026] S2 is specifically node feature extraction. There are millions of voxels in a 3D CT image. Therefore, it is necessary to express the connectivity in a sparse manner from the dense voxels. For this purpose, a 3D ground truth with dimensions D×H×W is divided into dimensions s d ×s h ×s wFor non-overlapping sub-regions, within each sub-region, the average position of the voxels belonging to the blood vessels is sampled as nodes. For sub-regions without coronary arteries, the central voxel is selected, and then a set of 3D nodes V is generated, which can be defined as:

[0027] and N n =[D / s d ×[H / s h ×[W / s w , where [.] means rounding the input to the smallest integer.

[0028] The connectivity constraint is the core step to characterize the blood vessel connectivity. For each pair of sampled nodes, if they are both located in the blood vessel region and there is a straight blood vessel channel between them, an edge should be constructed between them. However, for two nodes with a small Euclidean distance but belonging to two different blood vessel branches, they should not be connected by an edge. To handle this complex situation, the problem of whether there is a three-dimensional blood vessel path between nodes is transformed into the evaluation of the travel time between nodes. Specifically, for node V i ={x i , y i , z i}, this problem can be described by the following Eikonal equation: , where F(·) is the velocity function and T(·) is the travel time to be solved. Using the ground truth of the coronary artery as the velocity function, the voxels belonging to the blood vessels have a faster velocity than other voxels. Then, the fast marching method

[38] is used to solve the travel time from each node to vi. After that, if the travel time is less than the given threshold T th , an edge is constructed between the two nodes to obtain a three-dimensional edge set It should be noted that t ij <t th , where Eij and tij represent the edge and travel time between nodes vi and vj, respectively.

[0029] The connectivity prediction of the images in S3 is to combine two losses and retrain. Specifically, the connectivity of the graph network is trained. The three-dimensional blood vessel segmentation results of the patches are compared with the corresponding blood vessel annotations in CoronarySet2, and the training weight values and random parameter values are adjusted through error backpropagation training. The loss function is as follows:

[0030] where y i and For predicting the probability and the true situation of the nodes, N is the total number of voxels, the training weight value, which is the mapping relationship between the input and output during training and can be used to calculate the output value of each neuron. During the training process, the training weight value is the parameter that the model needs to optimize. Finally, in the embodiments of the present application, a specially constructed loss function is used, and this loss function is Loss seg and Loss cc a linear combination of and a specific constant:

[0031] Loss Total = α * Loss seg + β * Loss cc + λ * ||ε|| 2 2 , where ε is the trainable parameter of the entire network, and α, β, and λ are weight parameters. When training the three-dimensional visual graph network medical segmentation network, parameter optimization can be directly performed based on this training weight value, which will greatly reduce the training loss and improve the training efficiency, and obtain the fine segmentation result based on S3.

[0032] In S4, for the stitching of patches, the overall segmentation result of a single CTA can be obtained.

[0033] Specifically, for obtaining the three-dimensional coronary artery image data and constructing the coronary artery image dataset, the specific analysis is as follows: For different individuals, multiple three-dimensional coronary artery angiography images are collected using a computed tomography device, and the multiple three-dimensional coronary artery angiography images are stored in a standard medical image format to obtain the three-dimensional coronary artery image data; according to the resolution, size, and image quality of the three-dimensional coronary artery image data, preprocessing is performed on the three-dimensional coronary artery image data, and the preprocessing includes noise removal and image normalization operations; a coronary artery image dataset for training and testing is constructed, and the coronary artery image dataset includes the three-dimensional coronary artery image data of multiple different individuals, covering various anatomical variations of the coronary artery.

[0034] In this implementation, the specific analysis of extracting multi-scale texture features of coronary artery images based on the coronary artery image dataset and extracting the topological structure features of blood vessels based on the coronary artery topological feature extraction module is as follows: The coronary artery image dataset is input into the network, and then through the windowed self-attention mechanism and convolutional operations in Swin Transformer, feature information at different levels is extracted layer by layer. At different levels of the encoder, by fusing low-level and high-level features, richer texture information is extracted to obtain multi-scale texture features of coronary artery images. The multi-scale texture features of coronary artery images include local texture features and global texture features; a graph structure is constructed according to the blood vessel network of the coronary artery, with each blood vessel branch point as a node in the graph and the connection method of the blood vessels as the edges in the graph; according to the spatial distribution of the coronary artery blood vessels, the blood vessel branch nodes are automatically located, and the edges of the graph are formed through the connection relationship; a graph convolutional network is used to perform feature propagation on each node and edge in the graph to extract the topological structure features of the blood vessels. The topological structure features of the blood vessels include the connectivity of the blood vessels, the depth of the branches, the topological type of the blood vessels, the blood flow direction of each branch, the density of the blood vessel network, and the structural stability.

[0035] The three-dimensional image data of coronary CT angiography collected by a computed tomography (CT) device constitutes the dataset CoronarySet. For different individuals, multiple three-dimensional coronary angiography images are collected. The length and width of the image sizes of different individuals are the same, while the height is different. CoronarySet is split into the training sets CoronarySet1 and CoronarySet2, and the test sets CoronarySet3 and CoronarySet4 according to a certain ratio. Radiology experts are responsible for annotating the blood vessel pixels in the training set CoronarySet1 to obtain the blood vessel labels of the training set CoronarySet1.

[0036] Using the three-dimensional image dataset and multi-scale texture features can comprehensively and accurately describe the details of the coronary artery, helping doctors identify subtle lesions or abnormalities and enhancing the credibility of diagnosis; by collecting data for different individuals, personalized analysis of the coronary arteries of different patients can be provided. Considering anatomical variations, it can better adapt to the specific physiological conditions of patients; combining texture features and topological structures, more comprehensive information can be extracted through deep learning and graph convolutional network (GCN), integrating local and global information, which helps to discover subtle structural features that are difficult to capture by traditional methods; automated image data processing, feature extraction, and model training make data processing more efficient, reducing manual intervention and errors, which is particularly important for complex medical image analysis tasks; using standard medical image formats and common deep learning methods, it has strong scalability and can be applied to different medical data and devices, with good cross-platform adaptability.

[0037] Specifically, the specific analysis of fusing the multi-scale texture features of coronary artery images and the topological structure features of blood vessels and enhancing the correlation between different features through the attention mechanism is as follows: The attention mechanism is used to enhance the representation of important features and suppress unimportant features, thereby learning the correlation between the multi-scale texture features of coronary artery images and the topological structure features of blood vessels, and assigning appropriate weights to each feature; The feature fusion module takes the multi-scale texture features of coronary artery images and the topological structure features of blood vessels as inputs, and combines the weights of each feature assigned by the attention mechanism to form a comprehensive feature; The fused comprehensive feature vector is normalized and dimension-reduced, and then transmitted to the subsequent learning model.

[0038] In this implementation scheme, the texture features of the coronary artery region are specifically extracted by using multi-scale analysis methods (such as wavelet transform, Gabor filtering, etc.); the topological structure features of blood vessels are specifically extracted by using methods such as graph theory and topological analysis.

[0039] The specific logical steps of introducing the attention mechanism to weight the multi-scale texture features and topological structure features are as follows: First, extract the multi-scale texture features and topological structure features of the input image through a convolutional neural network (CNN) or other deep learning models; Calculate the similarity between each element of the input feature map (usually in the form of a dot product) to obtain an attention weight matrix representing the relationship between different positions; For each position, obtain a new representation through weighted average, use the global information of the feature map, and perform transformation through Query, Key, and Value to obtain the importance weight of each position; Apply the attention weight to the feature map to weight the features at each position, enhance the expression of key features, and suppress redundant or irrelevant features. Channel attention mechanism can also be adopted: Through pooling operations on the input feature map along the spatial dimensions (width and height), calculate the global feature representation of each channel, and these global features are mapped to the weights of each channel through a fully connected layer (or convolutional layer); Use activation functions such as sigmoid or softmax to normalize the weights, so that the weights of each channel are between 0 and 1, representing the importance of the features of that channel; Apply the calculated channel weights to each channel to obtain the weighted feature map.

[0040] The specific ways to obtain the comprehensive features include: direct splicing: splicing the multi-scale texture features and the vascular topological structure features as the input features, and the spliced feature vectors can be fed into the subsequent deep learning model for training; weighted fusion: calculating the weights of each feature and performing weighted fusion. Traditional machine learning methods (such as PCA, LDA) can be used to calculate the importance of each feature, or a neural network can be used to learn the weight of each feature; feature selection and dimensionality reduction: selecting the multi-scale texture features and topological structure features through feature selection algorithms (such as algorithms based on information gain and mutual information), reducing redundant features and retaining the most important information; multi-branch network: designing a multi-branch network, where each branch processes different types of features (such as the texture feature branch and the topological structure feature branch), and then fusing the outputs of each branch through weighted or splicing methods.

[0041] By using the attention mechanism, higher weights can be assigned to important features and lower weights to unimportant features, which helps to better capture and emphasize the internal relationship between the key texture features of coronary artery effects and the topological structure features of blood vessels; by introducing the attention mechanism in the feature fusion process, the model can adaptively optimize the contribution of each feature, making full use of the correlation between different features, thereby improving the accuracy and reliability of classification or regression tasks; using multi-scale texture features can capture different levels of detail of the coronary artery, while the topological structure features of blood vessels provide global information of the vascular network. After the two are combined, the model can comprehensively understand the morphological features and texture changes of blood vessels, which is beneficial to the early detection and analysis of diseases; the introduction of the attention mechanism can reduce the dependence on manually selected features and automatically evaluate the importance of each feature, making the model more flexible and efficient in complex data environments; the comprehensive feature vector after feature fusion is processed by normalization and dimensionality reduction, reducing redundant information and improving the training efficiency and stability of the subsequent learning model.

[0042] Specifically, the encoder is used to segment the multi-scale texture features and topological structure features of the fused coronary artery images, and then extract multi-scale features and reduce the resolution. The specific analysis is as follows: extract the fused comprehensive features, and then segment the fused comprehensive features into multiple small blocks based on the segmentation requirements. Each small block contains multi-scale texture features and topological features; use a convolutional neural network for convolutional operations to map the features in each small block to a higher-level feature space; reduce the spatial resolution of the features through pooling operations while retaining important information. At the same time, the encoder extracts the hierarchical features in the comprehensive features and reduces the resolution through step-by-step convolutional and pooling operations; fuse the features of different scales to form a multi-scale comprehensive feature map, and then use dimensionality reduction techniques to reduce the complexity of the high-dimensional features and map the high-dimensional features to a low-dimensional space to form low-resolution fused multi-scale features.

[0043] In this implementation scheme, the decoder is used to gradually restore the resolution to obtain the segmentation result, and then calculate the error between the segmentation result and the ground truth label, and optimize the network weights based on the error. The specific analysis is as follows: receive the low-resolution fused multi-scale features and their corresponding ground truth labels. The ground truth label is specifically the actual segmentation result of the image, which is used to train and optimize the network; gradually increase the spatial dimension of the low-resolution fused multi-scale features through upsampling or deconvolution operations until the low-resolution fused multi-scale features are restored to the image size of the original resolution, and obtain the intermediate results output at each step during the process of gradually increasing the spatial dimension. The intermediate results are specifically the gradually restored segmentation prediction maps; use the loss function to compare the intermediate results output at each step with the ground truth label in real time to obtain the loss value between the intermediate results output at each step and the ground truth label. The loss value reflects the difference between the network output and the real segmentation label; use the backpropagation algorithm to transmit the loss value back to each layer of the network, calculate the gradients of each layer, and then update the weight parameters of the network based on the calculated gradients. Specifically, update the weight parameters of the network based on the minimization of the loss value.

[0044] The loss value is used to measure the gap between the model output and the ground truth label. Commonly used loss functions include cross-entropy loss, Dice loss, etc. Through the optimization of the loss function, the network can effectively adjust the parameters and improve the segmentation accuracy; the backpropagation algorithm is used to calculate the gradients of each layer in the network and update the network weights according to the gradients, thereby reducing the value of the loss function. The gradients are calculated through the chain rule and represent the influence of each parameter on the final loss.

[0045] The data of the training set CoronarySet1 is used to train and test the Swin-UNetR segmentation model. The first part is sliced into multiple patches, and then the patches are input into the multi-layer Swin-UNetR encoder. After being processed by the max pooling layer, the size of the feature maps of the patches input to the encoder becomes smaller and the number of channels becomes larger. Then, through concatenation and transposed convolution layer processing, a feature map with the same size as the original input is obtained. Finally, through convolution operations and the softmax function, the predicted values of the CT images are obtained. The specific features of this encoder are as follows: First, the input CT images are divided into fixed-size image patches (Patches) by PatchPartition, and multi-scale features are extracted by the encoder. The Swin Transformer Block in the encoder efficiently captures local and global information using the sliding window attention mechanism. At the same time, the Patch Merging module gradually reduces the feature resolution and increases the number of channels. Finally, deep global features are extracted through the bottleneck layer. Then, the decoder gradually restores the resolution through the Patch Expanding module and combines the skip connections of the encoder to fuse multi-scale features. Each layer of the decoder uses the Swin Transformer Block to refine the features to ensure the feature expression ability, and finally outputs a segmentation map with the same size as the input, representing the category of each pixel. On this basis, the Swin-UNetR segmentation model is trained based on the following loss function, and after the model is stable, the segmentation loss is calculated through the test set CoronarySet3:

[0046] Through slicing and convolution operations, it is possible to capture features of different scales and levels in the image, so as to comprehensively understand the texture features and topological structure of the coronary artery at multiple scales; through pooling operations, the spatial resolution is reduced while important feature information is retained, avoiding the computational complexity problems brought by high-dimensional data; the decoder gradually restores the resolution through upsampling, thereby improving the prediction accuracy of the model. The network weights are optimized through backpropagation to further improve the segmentation accuracy; step-by-step operations (such as convolution, pooling, upsampling) make each link highly interpretable, and it is possible to analyze the feature extraction and changes at each step; the model can adapt to different types of input images and resolutions, adapt to complex coronary artery structures, and process image features of different scales.

[0047] Specifically, apply the trained segmentation network to the test data to obtain the three-dimensional segmentation result of the coronary artery. By comparing the gap between the predicted value and the true value, evaluate the accuracy of the segmentation result. The specific analysis of the connection constraint loss is as follows: Load and obtain the network weight parameters updated during the process of the decoder gradually restoring the resolution; Obtain the test data in the coronary artery image dataset, where the test data includes images and corresponding true labels; Perform normalization processing on the test images and adjust them to the input size used during training; Input the adjusted test images into the trained segmentation network for forward propagation to obtain the segmentation prediction result of the coronary artery; Recombine the segmentation prediction results of each slice to obtain the complete three-dimensional segmentation result, and evaluate the segmentation performance of the model based on the difference between the predicted segmentation result and the true label; Detect whether there are discontinuous regions in the three-dimensional segmentation result based on the connected regions in the three-dimensional segmentation result, that is, whether there is a connection loss situation; If a connection loss situation is detected, use morphological dilation operations or graph-based optimization methods to restore the lost regions.

[0048] In this implementation plan, by normalizing the test images and adjusting them to the input size used during training, the consistency of the input data is ensured, thereby reducing the impact of data preprocessing errors on the segmentation result; Using the trained segmentation network for forward propagation can make full use of the features and patterns learned during the training process, thereby improving the accuracy of the predicted segmentation result; By evaluating the difference between the prediction result and the true label, the performance of the model can be monitored in real time, and the model can be further adjusted and optimized; Recombining the segmentation prediction results of each slice to obtain the complete three-dimensional segmentation result can not only provide the segmentation result of the two-dimensional slice, but also analyze the three-dimensional structure of the entire coronary artery from a global perspective, enhancing the model's understanding of spatial relationships and making the segmentation result more accurate; The three-dimensional segmentation result makes it possible to detect subtle anatomical structures and complex morphologies, which is crucial for medical image analysis; For the problem of connection constraint loss in the test results, morphological dilation operations or graph optimization methods are used in the design to restore the lost connection regions. The morphological dilation operation can effectively fill the holes that may appear in the segmentation, the parts with connection loss, reducing the segmentation missing caused by noise or model errors; The graph-based optimization method can perform constraint optimization on a global scale, further ensuring the coherence of the segmentation result and avoiding the inconsistency that may be brought by local optimization; Connected region detection can detect whether there are discontinuous regions in the three-dimensional segmentation, which is particularly important for coronary artery segmentation because the vascular structure of the coronary artery usually needs to maintain continuity, and any connection loss may lead to inaccurate segmentation results. Through connected region analysis, not only the quality of the segmentation is improved, but also the model is further improved, especially in regions with complex morphologies or textures.

[0049] In summary, this application has at least the following effects:

[0050] Extracting multi-scale texture features of coronary artery images means that detailed information of blood vessels can be captured at different resolutions. Large blood vessels are extracted from low-resolution images, while thin blood vessels are extracted from high-resolution images, which increases the robustness and accuracy of processing. Through the coronary artery topology feature extraction module, the topological structure features of blood vessels can be obtained, which is crucial for maintaining the connectivity and integrity of blood vessels, especially when dealing with complex blood vessel networks. Fusing multi-scale texture features and topological structure features and enhancing the correlation between different features through the attention mechanism helps improve the accuracy of segmentation. The attention mechanism enables the network to focus more on important features and thus ignore irrelevant background information. Using the encoder to segment and reduce the resolution of the fused features and then gradually restoring the resolution through the decoder helps extract multi-scale features and maintain high-resolution detail information in the final segmentation result. By calculating the error between the segmentation result and the ground truth label and optimizing the network weights based on the error, the accuracy of segmentation can be continuously improved. In addition, by comparing the gap between the predicted value and the true value, the accuracy of the segmentation result can be evaluated, and information such as loss of connection constraints can be obtained, which is of great significance for further optimizing the network and improving the segmentation performance. Achieving three-dimensional segmentation of coronary arteries is of great significance for the early prevention and diagnosis of coronary artery diseases. Precise blood vessel segmentation results can provide valuable reference information for doctors, thus enabling the formulation of more effective treatment plans.

[0051] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0052] The present invention is described with reference to the flowchart of the method according to the embodiments of the present invention. It should be understood that the combination of each process in the flowchart can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for realizing the functions specified in Figure 1 one process or multiple processes.

[0053] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means embodying the function specified in one or more of the procedures Figure 1 or functions.

[0054] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the function specified in one or more of the procedures Figure 1 or functions.

[0055] Although the preferred embodiments of the present invention have been described, additional changes and modifications can be made by those skilled in the art once they learn of the basic inventive concept. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications falling within the scope of the present invention.

[0056] It is apparent that those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A method for continuous segmentation of blood vessels based on graph networks, characterized in that: The following steps are involved: Acquire coronary artery three-dimensional image data and construct a coronary artery image data set; The multi-scale texture features of coronary artery influence are extracted based on the coronary artery image dataset, and the topological structure features of blood vessels are extracted based on the coronary artery topological feature extraction module; The multi-scale texture features based on coronary artery influence and the topological structure features of blood vessels are fused, and the correlation between different features is enhanced through the attention mechanism; The encoder is used to segment the fused multi-scale texture features of the coronary artery and the topological structure features of the blood vessels, thereby extracting multi-scale features and reducing the resolution; The decoder is used to gradually restore the resolution to obtain the segmentation result, and then the error is calculated using the segmentation result and the true label, and the network weight is optimized based on the error; The trained segmentation network is applied to the test data to obtain the three-dimensional segmentation results of the coronary arteries. The accuracy of the segmentation results is evaluated by comparing the gap between the predicted values ​​and the true values, and the connection constraint loss is obtained.

2. The method for continuous segmentation of blood vessels based on graph network according to claim 1, characterized in that: The specific analysis of obtaining coronary artery three-dimensional image data and constructing a coronary artery image data set is as follows: A plurality of coronary artery three-dimensional angiography images are collected for different individuals using a computer tomography device, and the plurality of coronary artery three-dimensional angiography images are stored using a standard medical image format to obtain coronary artery three-dimensional imaging data; Preprocessing the coronary artery three-dimensional image data according to the resolution, size and image quality of the coronary artery three-dimensional image data, the preprocessing includes noise removal and image normalization operations; A coronary artery image dataset is constructed for training and testing respectively. The coronary artery image dataset includes coronary artery three-dimensional image data of multiple different individuals, covering a variety of anatomical variations of the coronary arteries.

3. The method for continuous segmentation of blood vessels based on graph network according to claim 1, characterized in that: The specific analysis of extracting the multi-scale texture features of coronary artery influence based on the coronary artery image dataset and extracting the topological structure features of blood vessels based on the coronary artery topological feature extraction module is as follows: The coronary artery image dataset is input into the network, and then the feature information of different levels is extracted layer by layer through the windowed self-attention mechanism and convolution operation in the Swin Transformer. At different levels of the encoder, richer texture information is extracted by fusing low-level and high-level features to obtain multi-scale texture features affected by the coronary artery, and the multi-scale texture features affected by the coronary artery include local texture features and global texture features; A graph structure is constructed based on the vascular network of the coronary arteries, with each vascular branch point as a node in the graph and the connection mode of the blood vessels as an edge in the graph; According to the spatial distribution of coronary arteries, the vascular branch nodes are automatically located and the edges of the graph are formed through connection relationships; A graph convolutional network is used to perform feature propagation on each node and edge in the graph to extract the topological structural features of the blood vessels, including the connectivity of the blood vessels, the depth of the branches, the topological type of the blood vessels, the blood flow direction of each branch, the density and structural stability of the vascular network.

4. The method for continuous segmentation of blood vessels based on graph network according to claim 1, characterized in that: The specific analysis based on the fusion of multi-scale texture features affected by coronary arteries and topological structure features of blood vessels and the enhancement of the correlation between different features through the attention mechanism is as follows: The attention mechanism is used to enhance the representation of important features and suppress unimportant features, thereby learning the correlation between the multi-scale texture features of coronary artery influence and the topological structure features of blood vessels, and assigning appropriate weights to each feature. The feature fusion module takes the multi-scale texture features of coronary artery influence and the topological structure features of blood vessels as input, and combines the weights of each feature assigned by the attention mechanism to form a comprehensive feature; The fused comprehensive feature vector is normalized and dimensionally reduced, and then passed to the subsequent learning model.

5. The method for continuous segmentation of blood vessels based on graph network according to claim 4, characterized in that: The encoder is used to segment the multi-scale texture features of the fused coronary artery and the topological structure features of the blood vessels, and then the multi-scale features are extracted and the specific analysis of reducing the resolution is as follows: Extract the fused comprehensive features, and then divide the fused comprehensive features into multiple small blocks based on the segmentation requirements. Each small block contains multi-scale texture features and topological features. Use convolutional neural networks to perform convolution operations to map the features in each small block to a higher-level feature space; The spatial resolution of the features is reduced through pooling operations while retaining important information. At the same time, the encoder extracts hierarchical features from the comprehensive features and reduces the resolution through step-by-step convolution and pooling operations; Features of different scales are fused to form a multi-scale comprehensive feature map, and then dimensionality reduction technology is used to reduce the complexity of high-dimensional features, mapping high-dimensional features to low-dimensional space to form low-resolution fused multi-scale features.

6. The method for continuous segmentation of blood vessels based on graph network according to claim 5, characterized in that: The decoder is used to gradually restore the resolution to obtain the segmentation result, and then the segmentation result and the true label are used to calculate the error. The specific analysis of optimizing the network weight based on the error is as follows: Receive the low-resolution fused multi-scale features and their corresponding true labels, where the true labels are specifically actual segmentation results of the images, for training and optimizing the network; The low-resolution fused multi-scale features are gradually increased in spatial dimension through upsampling or deconvolution operations until the low-resolution fused multi-scale features are restored to the image scale of the original resolution, and the intermediate results output at each step in the process of gradually increasing the spatial dimension are obtained. The intermediate results are specifically the gradually restored segmentation prediction images; The loss function is used in real time to compare the intermediate result output at each step with the true label to obtain the loss value between the intermediate result output at each step and the true label. The loss value reflects the difference between the network output and the true segmentation label. The loss value is passed back to each layer of the network through the back propagation algorithm, and the gradient of each layer is calculated. Then, based on the calculated gradient, the weight parameters of the network are updated using the optimization algorithm. The specific update of the weight parameters of the network is based on the minimization of the loss value.

7. The method for continuous segmentation of blood vessels based on graph network according to claim 1, characterized in that: The trained segmentation network is applied to the test data to obtain the 3D segmentation results of the coronary arteries. The accuracy of the segmentation results is evaluated by comparing the gap between the predicted values ​​and the true values. The specific analysis of the connection constraint loss is as follows: Load and obtain the updated network weight parameters during the decoder's gradual resolution recovery process; Obtain test data from the coronary artery image dataset, where the test data includes images and corresponding true labels; Normalize the test images and resize them to the input size used during training; Input the adjusted test image into the trained segmentation network and perform forward propagation to obtain the segmentation prediction result of the coronary artery; The segmentation prediction results of each slice are recombined to obtain a complete 3D segmentation result, and the segmentation performance of the model is evaluated based on the difference between the predicted segmentation result and the true label; Based on the connected regions in the three-dimensional segmentation results, it is detected whether there are discontinuous regions in the three-dimensional segmentation results, that is, whether there is a connection loss condition; If a connection loss is detected, the lost area is restored using morphological dilation operations or graph-based optimization methods.

Images