Coronary artery blood vessel image segmentation method and system, storage medium and electronic equipment

Abstract

The invention discloses a coronary artery blood vessel image segmentation method and system, a storage medium and electronic equipment, and the method comprises the steps: obtaining a to-be-processed coronary artery angiography X-ray image and a corresponding real label; shape self-adaptive convolution is initialized; constructing a shape adaptive convolutional network for coronary artery blood vessel segmentation; inputting the coronary artery blood vessel image into an encoder and a decoder to obtain a shared feature map; acquiring a coronary artery blood vessel prediction connected graph through a diagonal connectivity network according to the shared feature graph; obtaining a coronary artery segmentation mask through a segmentation network by the shared feature map; obtaining a coronary blood vessel prediction center line through the coronary blood vessel segmentation mask; a segmentation result is optimized through a hierarchical topology constraint module; by means of special convolution operation and hierarchical topology constraints, efficient learning of blood vessel features and topology is achieved, the limitation of a traditional method is overcome, the segmentation performance is improved, and coronary artery segmentation is accurately achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical data analysis, and particularly relates to a coronary artery vessel image segmentation method and system, a storage medium, and an electronic device. Background Art

[0002] The diagnosis of Coronary Artery Disease (CAD) depends on accurately segmenting the coronary arteries from coronary X-ray angiography images (CAG). Therefore, measuring the diameter of the narrowed arterial lumen is crucial for the quantitative assessment of coronary stenosis in clinical practice.

[0003] Traditional manual segmentation methods are time-consuming and vulnerable to the subjective factors of operators. In recent years, with the development of computer vision technology and machine learning algorithms, automated coronary artery segmentation techniques have been widely applied and developed. Such techniques mainly rely on deep learning models, such as Convolutional Neural Networks (CNNs), which can learn effective feature representations from a large amount of training data and then achieve efficient and accurate recognition and segmentation of the coronary artery region. However, the existing deep learning models for coronary artery vessel image segmentation still have the following problems:

[0004] Firstly, the complex structure of the coronary arteries, characterized by their meandering tree-like pipe morphology, poses a challenge to traditional square convolutions because it is difficult for traditional convolutions to effectively capture this complexity and extract meaningful features; secondly, the fine structures and small branches of the coronary arteries lead to a large number of mis-segmented regions, which are often broken or prone to under-segmentation.

[0005] In recent years, with the rapid development of deep learning technology, certain progress has been made in the research on coronary artery segmentation. For example, deep learning models such as the Fully Convolutional Network (FCN) and UNet have been applied to coronary artery segmentation tasks and shown good performance. However, these methods still have certain performance limitations when faced with the complex vascular shapes of the coronary arteries.

[0006] To address the above problems, some studies have proposed specific segmentation methods for tubular structures. For example, DSCNet achieves efficient learning of vascular features by applying one-dimensional convolutions in two directions; CoANet models the blood vessels based on road shape features using one-dimensional convolutions in four directions. These methods have improved the segmentation effect to a certain extent, but still do not fully consider the spatial distribution characteristics of the vascular structure, especially the non-strictly orthogonal nature of the vascular directions and the handling of a large number of fine branches. The existence of the above problems limits the effectiveness and universality of existing methods in practical applications. Summary of the Invention

[0007] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide a coronary artery vessel image segmentation method, system, storage medium and electronic device. This method can reduce the problems of loss of small blood vessel structures and small branches, and incorrect vascular network structures caused by the appearance of breakpoints, improve the feature expression ability of the model for tubular structure objects and the network reconstruction ability, enhance the segmentation performance, and accurately achieve coronary artery segmentation.

[0008] To achieve the above purpose, the present invention adopts the following technical solutions:

[0009] A coronary artery vessel image segmentation method includes the following steps:

[0010] Step 1: Obtain coronary artery vessel images, divide the training set, validation set and test set, and obtain the true labels, true connectivity graphs and true centerlines of the coronary artery vessels in the training set of coronary artery vessel image data.

[0011] Step 2: Design shape-adaptive convolution and construct a shape-adaptive convolution network.

[0012] The shape-adaptive convolution network consists of three parts: a feature extraction network, a diagonal connectivity network and a segmentation network.

[0013] Step 3: Take the training set of coronary artery vessel images as input, and use the feature extraction network to generate a shared feature map; take the shared feature map as input, and through the diagonal connectivity network, generate a predicted connectivity graph of the coronary artery vessels; generate a vascular distribution probability map through the segmentation network to obtain a coronary artery vessel segmentation mask; use the coronary artery vessel segmentation mask to extract the predicted centerline of the coronary artery vessels.

[0014] Step 4: Based on the true labels, true connectivity graphs and true centerlines of the coronary artery vessels in the training set of coronary artery vessel image data, and the predicted connectivity graph and predicted centerline of the coronary artery vessels obtained by taking the training set as input, combined with the hierarchical topology constraint module, optimize the shape-adaptive convolution network model constructed in Step 2.

[0015] Take the validation set as input, set the training parameters, and obtain the best network model for coronary artery vessel image segmentation.

[0016] Step 5: Take the coronary artery vessel image to be tested as input, generate a segmentation mask of the coronary artery vessels through the best network model, and obtain the coronary artery vessel segmentation result through binarization.

[0017] Preferably, the feature extraction network includes an encoder, a deep feature fusion module and a decoder.

[0018] The encoder is used to extract features from the coronary artery vessel image to obtain coronary artery vessel features. It contains 5 convolutional blocks, and only the first two convolutional blocks contain shape-adaptive convolutions;

[0019] The deep feature fusion module captures multi-scale features through multiple dilated convolutions to enhance the model's perception ability of multi-scale context information;

[0020] The decoder is used to gradually upsample the coronary artery features extracted by the encoder and fuse the features from the encoder to generate a shared feature map. It consists of 4 convolutional blocks, and the last three convolutional blocks contain shape-adaptive convolutions.

[0021] Further, the shape-adaptive convolution is initialized by initializing two diagonal-direction bar convolutions. The morphologies of the two bar convolutions are perpendicular to each other and their midpoints intersect, and the length is 9.

[0022] Preferably, the diagonal connectivity network consists of two pairs of dilated convolutions and a channel attention layer;

[0023] The dilated convolutions respectively use convolution sums with dilation coefficients of 1 and 3 to predict vessel connectivity maps at different scales. The channel attention layer learns the importance of each channel through adaptive average pooling and a fully connected layer.

[0024] Preferably, the segmentation network first extracts vessel features through a 3*3 convolution and keeps the number of channels of the feature map unchanged, and then uses a 1*1 convolution to generate a vessel probability distribution map, and the number of output channels is 1.

[0025] Preferably, the vessel centerline is calculated by a skeletonization algorithm.

[0026] Preferably, the hierarchical topology constraint module includes centerline constraint, surface constraint, and diagonal connectivity constraint, which obtain the topological information of the coronary artery vessels from three levels and optimize the vessel segmentation mask in backpropagation.

[0027] The present invention also protects a coronary artery vessel image segmentation system, including a vessel image acquisition module and a shape-adaptive convolution network module;

[0028] The vessel image acquisition module is used to acquire coronary artery vessel images;

[0029] The shape-adaptive convolution network module includes a shape-adaptive convolution module, a feature extraction module, a vessel connectivity map acquisition module, a coronary artery vessel segmentation module, a vessel centerline extraction module, and a hierarchical topology constraint module;

[0030] The described shape-adaptive convolution module is initialized as bar-shaped convolutions in two diagonal directions. The forms of the two bar-shaped convolutions are perpendicular to each other and their midpoints intersect, with a length of 9.

[0031] The described feature extraction module extracts features from coronary artery vessel images through a feature extraction network to generate a shared feature map of coronary artery vessels.

[0032] The vessel connectivity graph acquisition module is used to obtain a vessel connectivity prediction graph through a diagonal connectivity network.

[0033] The described coronary artery vessel segmentation module is used to generate a coronary artery vessel segmentation mask through a segmentation network and obtain the coronary artery vessel segmentation result using a binarization method.

[0034] The vessel centerline extraction module is used to extract the vessel centerline through a skeletonization algorithm.

[0035] The described hierarchical topology constraint module includes centerline constraint, surface constraint, and diagonal connectivity constraint, obtains the topological information of coronary artery vessels from three levels respectively, and optimizes the vessel segmentation mask during backpropagation.

[0036] The present invention also protects a storage medium that stores a program for implementing the coronary artery vessel image segmentation method as described above.

[0037] The present invention also protects an electronic device, including at least one central processing unit, a graphics processing unit, an input / output device, a network communication module, and a storage unit electrically coupled to the central processing unit; the storage unit stores program code that can be read and executed by the central processing unit to implement the coronary artery vessel image segmentation method as described above.

[0038] Compared with the prior art, the present invention has the following technical effects:

[0039] The coronary artery vessel image segmentation method of the present invention designs shape-adaptive convolution to construct a shape-adaptive convolution network for learning the structural features of coronary vessels. The shape-adaptive convolution network learns coronary vessels by using two shape-adaptive convolutions in diagonal directions to obtain an efficient coronary artery vessel feature map, providing a rich feature representation for subsequent segmentation and connected graph prediction. Then, a hierarchical topological constraint module is designed to improve the topological consistency of coronary vessels by learning the topological structures of lines, surfaces, and volumes of coronary vessels. That is, by adding centerlines, masks, and connected graphs, the problem that the topological structure error in coronary vessel segmentation cannot reconstruct a complete coronary structure can be alleviated, reducing the problems of loss of small vascular structures and small branches and incorrect vascular network structures caused by the appearance of breakpoints, improving the feature expression ability and network reconstruction ability of the model for tubular structure objects, enhancing the segmentation performance, and accurately realizing coronary artery segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 FIG. is a schematic flow chart of the construction of a shape-adaptive convolution network for coronary artery vessel image segmentation;

[0041] Figure 2 FIG. is a framework diagram of a shape-adaptive convolution network for coronary artery vessel image segmentation;

[0042] Figure 3 FIG. is a schematic diagram of the results of coronary artery vessel segmentation by different coronary artery vessel methods;

[0043] Figure 4 FIG. shows a schematic structural diagram of the electronic device in Embodiment 4 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The following further elaborates on the specific content of the present invention in conjunction with embodiments.

[0045] Embodiment 1

[0046] To solve the technical problems existing in the prior art, this embodiment provides a coronary artery vessel image segmentation method. First, a shape-adaptive convolution network is constructed and trained. Figure 1 FIG. is a schematic flow chart of the construction of a shape-adaptive convolution network for coronary artery vessel image segmentation, as Figure 1 shown, including the following steps:

[0047] Step 1: Obtain coronary artery vessel images. Specifically, when obtaining coronary artery vessel images, the patient is injected with an iodine-containing contrast agent, which can be visualized under X-rays and helps to clearly show the vascular structure. Then, the doctor inserts a slender catheter into the patient's artery (usually the artery in the thigh or wrist) and carefully guides it to the coronary artery under X-ray fluoroscopy. When the catheter reaches the appropriate position, the contrast agent is injected through the catheter, and at the same time, the X-ray machine is used to take rapid consecutive shots to obtain coronary artery angiography X-ray images.

[0048] Divide the obtained coronary artery vessel images into a training set, a validation set, and a test set, and obtain the true labels, true connectivity graphs, and true centerlines of the coronary artery vessels for the training set of coronary artery vessel image data;

[0049] Specifically, when obtaining the corresponding true vessel labels, professional radiologists or cardiovascular experts manually annotate the vessel regions on the coronary artery vessel images. When the experts annotate, they use dedicated medical image annotation software (such as ITK-SNAP, 3D Slicer, or Labelbox) to accurately annotate the vessel regions by drawing polygons, curves, or region filling. The true connectivity graph of the vessels can be represented as an n*n*4 matrix, where n represents the size of the true vessel label and 4 represents the four directions on the diagonal. For each vessel pixel point on the true vessel label, calculate whether it is connected to the neighboring pixel points in the four directions. If the pixel value of the neighboring point is 0, it means not connected; if the pixel value of the neighboring point is 1, it means connected. When obtaining the true centerline of the vessels, image erosion and dilation operations are used to refine the vessel boundaries and remove small branches and noise. Next, the width of the vessels is gradually reduced through a skeletonization algorithm (the Thinning algorithm in Opencv) until only a single-pixel-wide centerline remains. Finally, post-processing such as smoothing and optimization is performed on the extracted centerline to obtain the true centerline of the angiography image.

[0050] Step 2: Initialize the shape-adaptive convolution module and construct the shape-adaptive convolution network;

[0051] Specifically, the shape-adaptive convolution is initialized by initializing two bar convolutions in the diagonal directions. The lengths of the two initialized bar convolutions are 9. The morphologies of the two initialized bar convolutions should be perpendicular to each other and intersect at the midpoints. During network training, the two bar convolutions will adaptively adjust according to the structure of blood vessels, and this adjustment is achieved by using model parameters as the offsets of the two bar convolutions. For a bar convolution with a length of 9, 18 offsets will be generated, namely 9 horizontal offsets and 9 vertical offsets. The shape-adaptive convolution is used to replace the ordinary convolution, and normalization and activation functions are used afterwards. The BatchNorm method is used for normalization, and the ReLU method is used for the activation function.

[0052] Specifically, for the feature map X, assume Y 1 represents the convolution output in the positive diagonal direction, and Y 2 represents the convolution output in the anti-diagonal direction, and W represents the convolution kernel weight. Then, Y 1 and Y 2 can be respectively expressed as:

[0053]

[0054] where, Δx m and Δy m respectively represent the displacements in the x direction and y direction. For each pixel point m, there is a unique displacement. The final convolution result is the sum of Y 1 and Y 2 :

[0055] Y(i,j) = +Y 1 (i,j) + Y 2 (i,j)

[0056] As Figure 2 shown, the shape-adaptive convolution network includes a feature extraction network, a diagonal connectivity network, a segmentation network, as well as a blood vessel centerline extraction module and a hierarchical topology constraint module for optimizing the coronary artery blood vessel segmentation mask;

[0057] Among them, the feature extractor consists of an encoder, a deep feature fusion module and a decoder. There are 5 blocks in the encoder and 4 blocks in the decoder. Different dilated convolutions are used in the deep feature fusion module. For the use of the shape-adaptive convolution, the shape-adaptive network will use the shape-adaptive convolution in the first two blocks of the encoder and the last three blocks of the decoder.

[0058] The diagonal connectivity network consists of two pairs of dilated convolutions and channel attention layers. Each pair of dilated convolutions and channel attention layers generates a predicted vascular connectivity map. The difference is that different-sized dilated convolutions are used for the dilated convolutions and channel attention layers respectively to predict vascular connectivity maps at different scales. Dilated convolutions with sizes of 1 and 3 are used respectively.

[0059] The segmentation network first further extracts vascular features through a 3*3 convolution and keeps the number of channels of the feature map unchanged. Then a 1*1 convolution is used to generate a vascular probability distribution map. Preferably, the number of output channels is 1.

[0060] Step 3: ① Taking the coronary artery vascular images in the training set as input, the feature extraction network extracts features from the coronary artery vascular images to generate a shared feature map of the coronary artery vessels;

[0061] Specifically, the coronary artery vascular images are input into the first block of the encoder. The number of input channels is 3. After passing through 5 blocks in the encoder, the obtained high-order feature map is input into the depth feature fusion module to obtain multi-scale features. In the decoder part, the output of the corresponding block in the encoder is spliced into the block of the decoder through residual connection, and the shared feature map is obtained after passing through 4 blocks of the decoder.

[0062] ② Using the diagonal connectivity network to perform connectivity prediction on the coronary artery vascular pixel points and their diagonal adjacent points with the shared feature map, predicting whether the neighbor pixel points in the four directions around the vascular points are vascular regions, generating a connectivity probability map, and further obtaining a coronary artery vascular connectivity prediction map;

[0063] When the diagonal connectivity network is being trained, the true coronary artery vascular connectivity map is used as a label for training; and the true coronary artery vascular connectivity map is generated from the true vascular label.

[0064] When the diagonal connectivity network is being trained, dilated convolutions are used to generate connectivity maps with different intervals.

[0065] ③ Using the segmentation network to generate a coronary artery vascular segmentation prediction mask to obtain the coronary artery vascular segmentation result.

[0066] The segmentation network takes the shared feature map as input, obtains the probability distribution map of the coronary artery vessels, representing the probability that each pixel point is a vascular pixel, and obtains the coronary artery vascular segmentation prediction mask. When the segmentation network is being trained, the true coronary artery vascular label is used for training.

[0067] ④ Using the coronary artery vascular segmentation mask to extract the vascular centerline, obtaining a coronary artery vascular prediction centerline consistent with the coronary artery vascular segmentation mask;

[0068] Specifically, similar to the method for calculating the true centerline of blood vessels, image erosion and dilation operations are used to refine the blood vessel boundary, removing small branches and noise. Next, the width of the blood vessels is gradually reduced through a skeletonization algorithm until only a single-pixel-wide centerline remains. Finally, post-processing such as smoothing and optimization is performed on the extracted centerline to obtain the predicted centerline of the blood vessels in the angiography image.

[0069] Step 4: Based on the true labels of coronary arteries, the true connectivity graph of coronary arteries, and the true centerline of coronary arteries, and the predicted connectivity graph of coronary arteries, the predicted centerline of coronary arteries, and the predicted mask of coronary arteries obtained by using the training set as input, combined with the hierarchical topology constraint module, optimize the coronary artery segmentation mask to obtain the coronary artery segmentation result.

[0070] Specifically, the hierarchical topology constraint module includes centerline constraint, surface constraint, and diagonal connectivity constraint, obtaining the topological information of coronary arteries from three levels respectively, and optimizing the blood vessel segmentation mask during backpropagation.

[0071] The centerline constraint mainly uses the centerline of the segmentation result to constrain the continuity and topological correctness of the blood vessels. Based on the true label M of the blood vessels gt and the predicted blood vessel mask M pred calculate the true centerline L gt and the predicted centerline L pred .

[0072] Then, calculate the Dice loss between L gt and L pred as part of the overall model loss function to enhance the continuity of the segmentation result, defined as follows:

[0073]

[0074] The surface constraint is mainly to enhance the continuity of coronary artery segmentation. The Dice coefficient widely used in image segmentation tasks is introduced as a surface constraint into the hierarchical topology constraint. This coefficient evaluates the similarity by calculating the ratio of the overlapping part of two sets to the total, defined as follows:

[0075]

[0076] where D dice represents the Dice coefficient of the predicted value and the label, M gt represents the true blood vessel label, M pred represents the blood vessel mask predicted by the network, and the two are of equal size.

[0077] The diagonal connectivity constraint is mainly used to obtain the cube connectivity graph of the coronary artery by constraining the connectivity relationship between preset adjacent vascular points in the coronary artery vascular characteristics. Further, for each pixel point, we design a diagonal connection prediction branch (DC), considering adjacent points in the diagonal directions: top - left, top - right, bottom - left, and bottom - right. Therefore, the diagonal connection constraint works in collaboration with the shape - adaptive convolution to more effectively capture vascular characteristics. Given a pixel point P i,j , the set of its neighboring diagonal points is defined as:

[0078] N i,j = {P i-1,j-1 , P i-1,j+1 , P i+1,j-1 , P i+1,j+1}

[0079] where P i-1,j-1 , P i-1,j+1 , P i+1,j-1 , P i+1,j+1 represent the neighbors in the top - left, top - right, bottom - left, and bottom - right directions of P i,j , respectively. Therefore, the diagonal connection C can be defined as follows:

[0080]

[0081] where N k represents the k - th neighbor in the set of neighboring points N i,j , and the value range of k is from 1 to 4. In the training stage, the DC branch generates a 4 - dimensional prediction vector for each pixel point, forming a 4 - dimensional probability map C pred for quantifying the connection probability, as shown in Figure 2 . Finally, the loss between the ground truth C gt and the predicted value C pred can be expressed as:

[0082]

[0083] where y i and represent the values at the corresponding positions in C gt and C pred , respectively.

[0084] It should be noted that in the training stage, the hierarchical topology constraint module uses the pre - calculated coronary artery vascular centerline and connectivity graph labels, plus the ground truth labels of the blood vessels, to jointly constrain and optimize the vascular segmentation mask. In the actual test stage, without the above three labels, the network directly predicts the segmentation result of the coronary artery blood vessels based on the vascular characteristics learned during the training process.

[0085] Furthermore, the loss function in the training process of the hierarchical topology constraint module is expressed as:

[0086] L = A + λB + (1 - λ)C

[0087] Where L represents the loss function, A represents the loss value between the predicted result and the true result of the blood vessels, B represents the loss value between the predicted centerline of the blood vessels and the true centerline, C represents the loss value between the predicted connectivity graph of the blood vessels and the true connectivity graph, and λ represents a parameter used to balance the centerline loss value and the connectivity graph loss value.

[0088] The topological information of the coronary artery blood vessels is obtained from three levels respectively, and the blood vessel segmentation mask is optimized in backpropagation.

[0089] Further, using the validation set as the input and setting the training parameters, the best network model for coronary artery blood vessel image segmentation is obtained.

[0090] Step Five: After obtaining the best network model for coronary artery blood vessel image segmentation, perform coronary artery blood vessel image segmentation on the data to be tested, which specifically includes the following steps:

[0091] Step 1, obtain the coronary artery blood vessel image to be tested;

[0092] Specifically, the coronary artery blood vessel images in the test phase come from different patients or different medical image datasets; when obtaining the coronary artery blood vessel images, the patient is injected with an iodine-containing contrast agent, which can be visualized under X-ray to help clearly show the blood vessel structure. Then, the doctor inserts a slender catheter into the patient's artery (usually the artery in the thigh or wrist), and carefully guides it to the coronary artery under X-ray fluoroscopy. When the catheter reaches the appropriate position, the contrast agent is injected through the catheter, and at the same time, the X-ray machine is used to take rapid consecutive shots to obtain coronary artery angiography X-ray images.

[0093] Step 2, perform coronary artery blood vessel segmentation through the best network model;

[0094] Specifically, input the coronary artery blood vessel image to be tested into the shape-adaptive convolutional network;

[0095] The image is first processed through multiple blocks of the encoder of the feature extraction network to extract high-order features, and multi-scale features are generated through the depth feature fusion module, and then enter the decoder part to obtain the final shared feature map;

[0096] The connectivity of the coronary artery blood vessel pixel points and their diagonal adjacent points in the shared feature map is predicted through the diagonal connectivity network to obtain the coronary artery blood vessel connectivity graph;

[0097] Generate a segmentation mask of coronary arteries through a segmentation network based on the shared feature map, which represents the probability that each pixel point is a blood vessel. According to the blood vessel mask output by the model, use the binarization method to obtain the segmentation result of coronary arteries. Each value in the probability map of the blood vessel mask represents the probability that the corresponding pixel point belongs to the blood vessel area, and the range is 0 to 1. For pixel points with a probability greater than 0.5, it is regarded as the blood vessel area; for pixel points less than or equal to 0.5, it is regarded as the background area.

[0098] To illustrate the accuracy of the coronary artery image segmentation method in the embodiments of the present invention, this embodiment compares the specific implementation data of the method of the present invention with the existing deep learning-based blood vessel segmentation methods. The existing deep learning-based blood vessel segmentation methods include CoANet, DSCNet, ResUNet, UNet, CS 2 -Net and clDice. The quantitative results are shown in Table 1, and the qualitative results are as Figure 3 shown.

[0099]

[0100] Table 1 summarizes the index results of seven different methods in the coronary artery segmentation task. It can be seen that the method proposed in this paper is significantly better than the methods proposed in other literatures. Specifically, our method performs better than other methods in various metrics. In particular, it achieves excellent Dice coefficients and intersection over union (IoU) scores of 81.83% and 70.11% respectively, exceeding the advantages of CoANet by 2.19% and 2.86% respectively. This is attributed to the fact that our method enhances the ability to accurately capture the complex features of the vascular tree-like tubular structure and ensures the topological consistency of the segmentation results, thereby improving the segmentation quality.

[0101] Figure 3 shows the schematic diagram of the segmentation results of four typical subjects in the first embodiment of the present invention using seven different coronary artery blood vessel methods; from Figure 3 the observed results, it can be seen that when the coronary arteries are affected by other types of blood vessels, all competing methods show missegmentation (for example, the blue parts in the first and second rows), while our method closely conforms to the ground truth. In addition, in the case where there are a large number of small branches inside the coronary arteries, most competing methods tend to show discontinuous segmentation and missegmentation (for example, the red parts in the last two rows). In contrast, our method always maintains high-quality topological correctness under the guidance of hierarchical topological constraints. The qualitative results further verify the effectiveness of the coronary artery segmentation method proposed in this paper.

[0102] The coronary artery vessel image segmentation method provided in this embodiment combines shape-adaptive convolution and hierarchical topological constraints. Experimental results show that significant improvements have been achieved in terms of overlap, accuracy, and topological structure compared with state-of-the-art methods. Our main contribution lies in introducing a novel convolution pattern, which is constructed based on traditional square convolution to better learn the spatial shape of coronary arteries. In addition, through specially designed hierarchical topological constraints, our method achieves more accurate segmentation results, which helps clinicians make accurate diagnoses.

[0103] Embodiment 2

[0104] To solve the above technical problems existing in the prior art, this embodiment provides a coronary artery vessel image segmentation system, including a vessel image acquisition module and a shape-adaptive convolution network module;

[0105] The vessel image acquisition module is used to acquire coronary artery vessel images;

[0106] The shape-adaptive convolution network module includes a shape-adaptive convolution module, a feature extraction module, a vessel connectivity graph acquisition module, a coronary artery vessel segmentation module, a vessel centerline extraction module, and a hierarchical topological constraint module;

[0107] The shape-adaptive convolution module is initialized as strip convolutions in two diagonal directions. The morphologies of the two strip convolutions are perpendicular to each other and their midpoints intersect, with a length of 9.

[0108] The feature extraction module performs feature extraction on the coronary artery vessel image through a feature extraction network to generate a shared feature map of the coronary artery vessels;

[0109] The vessel connectivity graph acquisition module is used to obtain a vessel connectivity prediction graph through a diagonal connectivity network;

[0110] The coronary artery vessel segmentation module is used to generate a coronary artery vessel segmentation mask through a segmentation network and obtain a coronary artery vessel segmentation result using a binarization method;

[0111] The vessel centerline extraction module is used to extract the vessel centerline through a skeletonization algorithm.

[0112] The hierarchical topological constraint module includes a centerline constraint, a surface constraint, and a diagonal connectivity constraint, which obtain the topological information of the coronary artery vessels from three levels respectively and optimize the vessel segmentation mask during backpropagation.

[0113] The coronary artery vessel image segmentation system of this embodiment constructs a shape-adaptive convolutional network for learning the structural features of coronary vessels through shape-adaptive convolution. This network learns coronary vessels by using two shape-adaptive convolutions in diagonal directions, obtaining an efficient coronary artery vessel feature map, which provides a rich feature representation for subsequent segmentation and connected graph prediction. Then, a hierarchical topological constraint module is designed to improve the topological consistency of coronary vessels by learning the topological structures of coronary vessels in lines, planes, and volumes. That is, by adding centerlines, masks, and connected graphs, the problem that the topological structure error in coronary artery vessel segmentation cannot reconstruct the complete coronary structure can be alleviated, the feature expression ability and topological reconstruction ability of the model for tubular structure objects can be improved, the segmentation performance can be enhanced, and the coronary artery segmentation is accurately realized.

[0114] Embodiment 3

[0115] This embodiment provides a storage medium on which a program for implementing the coronary artery segmentation method can be stored.

[0116] A computer-readable storage medium stores computer instructions, and when the instructions are executed by a computer, the following steps are implemented:

[0117] Image preprocessing: Obtain vascular image data from a medical imaging device, and perform denoising, enhancement, and normalization processing on the image to improve the input quality of the segmentation model.

[0118] Model loading and initialization: Load a deep learning-based vascular segmentation model and initialize the parameters of the model.

[0119] Segmentation inference: Input the preprocessed image data, extract features using the shape-adaptive convolutional network module, and generate a segmentation result in combination with the hierarchical topological constraint module.

[0120] Post-processing: Process the segmentation result output by the model, including removing small targets, smoothing edges, and morphological operations, to obtain a more accurate vascular segmentation map.

[0121] Result storage and display: Save the segmentation result as an image file or a data file for subsequent diagnosis or analysis, and perform visual display on the user interface.

[0122] The storage medium can be a hard disk, a solid-state drive, an optical disc, a flash drive, or a cloud storage device, etc., which can store the above instructions in electronic form for execution.

[0123] Embodiment 4

[0124] This embodiment provides an electronic device, which includes: at least one central processing unit (CPU), a graphics processing unit (GPU); a storage unit electrically coupled to the central processing unit (CPU); and an input / output device and a network communication module. The storage unit stores a program code that can be read and executed by the central processing unit (CPU) and is used to execute the processing method related to the present invention.

[0125] Figure 4 FIG. 4 shows a schematic diagram of the structure of an electronic device in Embodiment 4 of the present invention. Figure 4 , the electronic device of the present invention comprises:

[0126] Computing unit: The computing unit includes a central processing unit (CPU) and a graphics processing unit (GPU). The CPU has multi-core parallel processing capabilities, which is used to perform data preprocessing, coordination and scheduling of computing tasks, and supports efficient thread management to meet real-time requirements. The GPU significantly improves the efficiency of deep learning-related calculations through hardware acceleration, supports large-scale matrix operations and network reasoning capabilities of complex models, and its architecture should optimize parallel computing performance and adapt to deep learning frameworks and algorithm requirements. The two work together to efficiently process the large amount of data calculations and real-time result output involved in medical image analysis, thereby meeting high-performance and low-latency processing requirements.

[0127] Storage unit: The storage unit includes random access memory (RAM) and non-volatile memory (such as solid-state hard disk). RAM is used to store temporary data generated during operation and intermediate state data of the program. Its capacity and access speed are crucial to system performance. It should support high concurrent access to meet the needs of multi-threaded or multi-tasking processing, thereby ensuring the smoothness and real-time nature of the calculation process. Non-volatile memory is used to store program files, model parameters and long-term retained data. Solid-state hard disks (SSDs) or other non-volatile storage media can be used to provide fast data writing and reading capabilities. Its capacity design needs to support the storage and loading of high-resolution format files of massive medical images. In addition, the storage unit can adopt a hierarchical storage architecture, combining cache and main memory to optimize data access; be equipped with a data backup module to improve data reliability through mirror storage or RAID (redundant array of independent disks); and support the expansion of storage capacity through external interfaces (such as USB, SATA or PCIe) to meet the needs of future data growth. Through the above design, the storage unit can efficiently support the data processing needs of electronic devices and ensure the stability and efficiency of the system in high-performance computing and medical image processing.

[0128] Input device: The input device supports the acquisition of medical images, including but not limited to various medical imaging devices such as CT, MRI, angiography, etc., and can achieve the acquisition of high-resolution image data. The input device needs to be compatible with standard medical image formats (such as DICOM) to ensure the consistency of data interaction and processing between devices, and at the same time support multiple data input interfaces to adapt to the connection requirements of different acquisition devices. Through the input device, medical image data can be efficiently acquired and transmitted, providing a basis for subsequent processing and analysis.

[0129] Output device: The output device includes a high-resolution display for visually displaying the vascular segmentation results to ensure that the details of the results are clearly visible. In addition, the output device can also include an interactive visualization device (such as a device supporting 3D display) to enable in-depth analysis and interactive operations on the segmentation results. These devices can support multiple display modes to meet the needs of clinical diagnosis or scientific research analysis and provide users with a more comprehensive and intuitive image visualization effect.

[0130] Network communication module: It includes a high-speed Ethernet adapter and an optional 5G communication module for remote data transmission and edge computing scenarios.

[0131] Through the configuration of the above electronic devices, efficient data processing and real-time interactive analysis of medical image segmentation results of the present invention can be achieved. The device structure described in this embodiment is only an exemplary configuration, and the specific implementation scheme can be adjusted according to requirements.

[0132] Although the disclosed embodiments of the present invention are as above, the content described is only an embodiment adopted for the convenience of understanding the present invention and is not intended to limit the present invention. Any person skilled in the art within the technical field to which the present invention pertains, without departing from the spirit and scope disclosed by the present invention, can make any modifications and changes in the form and details of the implementation, but the protection scope of the present invention shall still be subject to the scope defined by the appended claims.

Claims

1. A coronary artery image segmentation method, characterized in that: The following steps are involved: Step 1: Obtain a coronary artery image, divide it into a training set, a validation set, and a test set, and obtain the coronary artery true label, the coronary artery true connectivity map, and the coronary artery true centerline of the coronary artery image data of the training set; Step 2: Design shape adaptive convolution and build a shape adaptive convolution network; The shape adaptive convolutional network consists of three parts: feature extraction network, diagonal connectivity network and segmentation network; Step 3: using the coronary artery image of the training set as input, using the feature extraction network to generate a shared feature map; using the shared feature map as input, using the diagonal connectivity network to generate a coronary artery prediction connectivity map; using the segmentation network to generate a vascular distribution probability map to obtain a coronary artery segmentation mask; using the coronary artery segmentation mask to extract the coronary artery prediction centerline; Step 4: Based on the coronary artery true labels, coronary artery true connectivity graph and coronary artery true centerline of the coronary artery image data of the training set, and the coronary artery segmentation mask, coronary artery predicted connectivity graph and coronary artery predicted centerline obtained with the training set as input, the shape adaptive convolutional network model constructed in step 2 is optimized in combination with the hierarchical topological constraint module; Using the validation set as input, the training parameters were set to obtain the best network model for coronary artery image segmentation; Step 5: Taking the coronary artery image to be tested as input, a segmentation mask of the coronary artery is generated through the optimal network model, and the coronary artery segmentation result is obtained through binarization.

2. The coronary artery image segmentation method according to claim 1, characterized in that: The feature extraction network includes an encoder, a deep feature fusion module and a decoder; The encoder is used to extract features from the coronary artery image to obtain coronary artery features, and includes 5 convolution blocks, and only the first two convolution blocks include shape adaptive convolution; The deep feature fusion module captures multi-scale features through multiple hole convolutions to enhance the model's perception of multi-scale contextual information; The decoder is used to gradually upsample the coronary artery features extracted by the encoder and fuse the features from the encoder to generate a shared feature map; it consists of 4 convolution blocks, and the last three convolution blocks contain shape adaptive convolutions.

3. The coronary artery image segmentation method according to claim 1 or 2, characterized in that: The shape adaptive convolution is initialized by initializing two diagonal strip convolutions, where the two strip convolutions are perpendicular to each other and intersect at their midpoints, and have a length of 9.

4. The coronary artery image segmentation method according to claim 1, characterized in that: The diagonal connectivity network consists of two pairs of dilated convolution and channel attention layers; The dilated convolution uses convolutions with dilation coefficients of 1 and 3 to predict vascular connectivity maps of different scales; the channel attention layer learns the importance of each channel through adaptive average pooling and a fully connected layer.

5. The coronary artery image segmentation method according to claim 1, characterized in that: The segmentation network first extracts the vascular features through a 3*3 convolution and keeps the number of channels of the feature map unchanged, and then uses a 1*1 convolution to generate a vascular probability distribution map, and the number of output channels is 1.

6. The coronary artery image segmentation method according to claim 1, characterized in that: The blood vessel centerline is calculated by a skeletonization algorithm.

7. The coronary artery image segmentation method according to claim 1, characterized in that: The hierarchical topology constraint module includes centerline constraint, surface constraint, and diagonal connectivity constraint, respectively obtaining topological information of coronary artery vessels from three levels, and optimizing vessel segmentation mask in back propagation.

8. A coronary artery image segmentation system, characterized in that: It includes a blood vessel image acquisition module and a shape adaptive convolutional network module; The blood vessel image acquisition module is used to acquire a coronary artery blood vessel image; The shape adaptive convolutional network module includes a shape adaptive convolution module, a feature extraction module, a vascular connectivity map acquisition module, a coronary artery segmentation module, a vascular centerline extraction module and a hierarchical topology constraint module; The shape adaptive convolution module is initialized as two diagonal strip convolutions, the two strip convolutions are perpendicular to each other and intersect at their midpoints, and have a length of 9. The feature extraction module extracts features from the coronary artery image through a feature extraction network to generate a shared feature map of the coronary artery; The vascular connectivity map acquisition module is used to acquire a vascular connectivity prediction map through a diagonal connectivity network; The coronary artery segmentation module is used to generate a coronary artery segmentation mask through a segmentation network and obtain a coronary artery segmentation result using a binarization method; The blood vessel centerline extraction module is used to extract the blood vessel centerline through a skeletonization algorithm. The hierarchical topology constraint module includes centerline constraint, surface constraint, and diagonal connectivity constraint, which respectively obtains the topological information of the coronary artery vessels from three levels and optimizes the vessel segmentation mask in the back propagation.

9. A storage medium, characterized in that: The device stores a program for implementing the coronary artery image segmentation method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: The invention comprises at least one central processing unit, a graphics processing unit, an input / output device, a network communication module, and a storage unit electrically coupled to the central processing unit; the storage unit stores a program code which can be read and executed by the central processing unit and implements the coronary artery vascular image segmentation method according to any one of claims 1 to 7.

Images