CNN and Transform-based three-dimensional coronary artery segmentation method and system

Abstract

The invention belongs to the technical field of medical image segmentation, and provides a three-dimensional coronary artery segmentation method and system based on a CNN and a Transform, and the method comprises the steps: constructing a neural network model comprising an encoder and a decoder based on the CNN and the Transform, and enabling a bottleneck layer in the neural network model to comprise a multi-scale feature interaction module; inputting a three-dimensional coronary artery image to be segmented into the neural network model; wherein the features of the three-dimensional coronary artery image to be segmented are extracted in the encoder, and the features are integrated at a bottleneck layer through the multi-scale feature interaction module; and carrying out weighted fusion on the integrated features and sampling features output by each layer of the decoder in the decoder, and outputting a segmentation result. According to the invention, the defect of fragmentation of the segmentation result in the prior art is overcome, and a more accurate segmentation method is provided.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image segmentation, and in particular relates to a three-dimensional coronary artery segmentation method and system based on CNN and Transformer. Background Art

[0002] Cardiovascular disease is the leading cause of harm to human health and death in China and around the world. Coronary computed tomography angiography (CCTA) is a common method for diagnosing coronary artery disease. CCTA can identify calcified and non-calcified plaques, which can help clinicians determine plaque characteristics and develop treatment plans. However, the location and severity of plaques that cause vascular stenosis in CCTA images currently require manual evaluation by radiologists, which is not only time-consuming but also prone to misdiagnosis and omission. The accurate interpretation and diagnosis of cardiovascular disease largely depends on the personal experience and expertise of clinicians. Therefore, automatic segmentation of coronary arteries can help doctors in subsequent clinical diagnosis and treatment.

[0003] At present, the existing methods of segmenting coronary arteries using neural networks mainly use the U-Net network model to automatically learn image features, and combine image preprocessing and data augmentation to segment coronary arteries. The preprocessing process has limited effect, resulting in insufficient optimization of image quality and consistency. In addition, since the segmentation process is based on small-sized image sub-blocks and a relatively simple network structure, the model can only learn local coronary artery structures and is difficult to fully capture the overall three-dimensional morphology of the coronary arteries, resulting in fragmented segmentation results, discontinuous coronary artery branches, and even fractures, which cannot accurately reflect the overall contour and complex characteristics of the coronary arteries. Summary of the invention

[0004] In order to overcome the defects of the above-mentioned prior art that the segmentation results are fragmented and cannot accurately reflect the overall contour and complex features of the coronary arteries, the present invention proposes a 3D coronary artery segmentation method and system based on CNN and Transformer.

[0005] In order to achieve the above technical effects, the technical solution of the present invention is as follows:

[0006] A 3D coronary artery segmentation method based on CNN and Transformer comprises the following steps:

[0007] A neural network model including an encoder and a decoder is constructed based on CNN and Transformer, wherein the bottleneck layer in the neural network model includes a multi-scale feature interaction module;

[0008] The three-dimensional coronary artery image to be segmented is input into the neural network model; wherein, the features of the three-dimensional coronary artery image to be segmented are extracted in the encoder, and the features are integrated through the multi-scale feature interaction module in the bottleneck layer; and in the decoder, the integrated features and the sampling features output by each layer of the decoder are weightedly fused and the segmentation result is output.

[0009] The present invention also proposes a 3D coronary artery segmentation system based on CNN and Transformer, the system comprising:

[0010] Model building unit: constructing a neural network model including an encoder and a decoder based on CNN and Transformer, wherein the bottleneck layer in the neural network model includes a multi-scale feature interaction module;

[0011] Image segmentation unit: inputting the three-dimensional coronary artery image to be segmented into the neural network model; wherein, in the encoder, extracting the features of the three-dimensional coronary artery image to be segmented, integrating the features through the multi-scale feature interaction module at the bottleneck layer; in the decoder, weighted fusion of the integrated features and the sampling features output by each layer of the decoder is performed and the segmentation result is output.

[0012] The present invention also proposes an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the three-dimensional coronary artery segmentation method based on CNN and Transformer as described in the present invention is implemented.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] The present invention can capture the complex morphological features of the coronary arteries at different scales through a multi-scale feature interaction module, significantly improving the integrity and accuracy of the segmentation results, thereby reducing the fragmentation phenomenon commonly seen in traditional segmentation methods, ensuring that the segmentation results are more complete and accurately reflect the overall contour and complex features of the coronary arteries. In addition, in the decoder part, the present invention further enhances the focus on the coronary artery position by weighted fusion of the integrated features and the upsampled features, effectively reducing background interference and improving the distinguishability of the coronary arteries from the background. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flowchart of a three-dimensional coronary artery segmentation method based on CNN and Transformer.

[0016] Figure 2 This is a structural diagram of the neural network model based on CNN and Transformer.

[0017] Figure 3 This is the structural diagram of the 3D encoding block.

[0018] Figure 4 Structural diagram of the attention-weighted feature fusion layer.

[0019] Figure 5 This is the structural diagram of the multi-scale feature interaction module.

[0020] Figure 6 This is an architecture diagram of a 3D coronary artery segmentation system based on CNN and Transformer.

[0021] Figure 7 is the actual segmentation result of a certain 3D coronary artery image.

[0022] Figure 8 This is the result of segmenting a three-dimensional coronary artery image using the three-dimensional coronary artery segmentation method based on CNN and Transformer proposed in this scheme. DETAILED DESCRIPTION

[0023] The accompanying drawings are only used for illustrative purposes and are not to be construed as limiting the present invention;

[0024] It is understandable to those skilled in the art that some well-known descriptions may be omitted in the drawings.

[0025] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0026] Example 1

[0027] This embodiment proposes a 3D coronary artery segmentation method based on CNN and Transformer. Figure 1 , which is a flow chart of a 3D coronary artery segmentation method based on CNN and Transformer in this embodiment.

[0028] The 3D coronary artery segmentation method based on CNN and Transformer proposed in this embodiment includes the following steps:

[0029] A neural network model including an encoder and a decoder is constructed based on CNN and Transformer, wherein the bottleneck layer in the neural network model includes a multi-scale feature interaction module;

[0030] The three-dimensional coronary artery image to be segmented is input into the neural network model; wherein, the features of the three-dimensional coronary artery image to be segmented are extracted in the encoder, and the features are integrated through the multi-scale feature interaction module in the bottleneck layer; and in the decoder, the integrated features and the sampling features output by each layer of the decoder are weightedly fused and the segmentation result is output.

[0031] In this embodiment, a neural network model based on CNN and Transformer is established, a multi-scale feature interaction module and an attention weighted feature fusion module are designed at the bottleneck layer, and the coronary artery image to be segmented is input into the neural network model to obtain the segmented coronary artery output. The multi-scale features of the coronary artery are extracted through the multi-scale feature interaction module, which solves the problem of incorrect segmentation of the coronary artery and large-area fragment segmentation; and the integrated features are weightedly fused with the up-sampled features in the decoder to focus on the position of the coronary artery, enhance the distinguishability between the coronary artery and the background, improve the accuracy and clarity of the segmentation, and help obtain higher quality coronary artery segmentation results.

[0032] In an optional embodiment, the three-dimensional coronary artery image to be segmented is randomly cut into n three-dimensional blocks of size a×a×a; and the three-dimensional blocks are input into the neural network model using a sliding window method.

[0033] In this embodiment, the computational complexity and memory usage are significantly reduced by dividing the image into small blocks, and the generalization ability of the model is enhanced and the risk of overfitting is reduced through random cropping; fixed-size blocks not only retain the local structural characteristics of the blood vessels, but also enable the model to focus more on details, which is conducive to the accurate segmentation of small blood vessel branches.

[0034] In an optional embodiment, the encoder and the decoder adopt a multi-layer symmetrical structure, the encoder includes a plurality of 3D encoding blocks connected in sequence; the decoder includes a plurality of decoding blocks connected in sequence.

[0035] like Figure 2 As shown in the figure, it is a structural diagram of the neural network model based on CNN and Transformer.

[0036] In this embodiment, the encoder is used to downsample the image to extract feature information, and the decoder is used to upsample the image to restore the original size and segment the blood vessel contour.

[0037] In an optional embodiment, the 3D encoding block includes a swin-transformer layer and a convolution block layer; the features in the 3D encoding block are divided into two features with the same number of channels and input into the swin-transformer layer and the convolution block layer respectively, and the outputs of the swin-transformer layer and the convolution block layer are spliced ​​as the output of the 3D encoding block.

[0038] like Figure 3 As shown, it is the structural diagram of the 3D encoding block.

[0039] In this embodiment, the swin-transformer layer can capture global features, which helps to extract the overall morphology of the coronary artery and improve the recognition ability of complex structures. The convolution block layer can capture local detail information, enhance the resolution of subtle features, and improve the accuracy of the segmentation results. In addition, several 3D encoding blocks can extract features of different depths layer by layer, so that the model can capture richer feature representations in a hierarchical manner, thereby improving the accuracy and stability of coronary artery segmentation.

[0040] As an exemplary illustration, the encoder includes four layers of 3D encoding blocks, namely, a first swin-transformer layer, a first convolutional block layer, a second swin-transformer layer, a second convolutional block layer, a third Sswin-transformer layer, a third convolutional block layer, a fourth swin-transformer layer and a fourth convolutional block layer, which are connected in sequence.

[0041] Specifically, the 3D encoding block first takes the input feature X∈R C×H×W×D , divided into two features with the same number of channels along the channel dimension, and then input into the swin-transformer layer and the convolution block layer respectively, and the outputs of these two modules are concatenated together as the output of the entire 3D encoding block.

[0042] The swin-transformer layer consists of four swin transformer blocks, alternating between the normal window attention mechanism and the attention mechanism with shifted windows. Two adjacent swin transformer blocks in the swin-transformer layer can be expressed by the following formula:

[0043]

[0044] Among them, 3DW-MSA() and 3DSW-MSA() represent the conventional window mechanism and the shifted window mechanism respectively. represents the output of the conventional window mechanism and the shifted window mechanism at layer l, y l is the output of MLP at layer l.

[0045] The convolutional block layer is composed of a 3×3×3 convolutional layer, a LayerNorm layer, and a Relu layer stacked twice in sequence. The calculation process of the convolutional block layer is as follows:

[0046] Y1 = Relu(LN(3DConv(X)))

[0047] Y2=Relu(LN(3DConv(Y1)))

[0048] Y=Y2+X

[0049] Among them, X represents the input of the convolution block layer, Y represents the output of the convolution block layer, Y1 and Y2 represent intermediate temporary variables, Relu() represents the activation function, 3DConv(X) represents three-dimensional depth-separable convolution, and LN() represents the normalization function.

[0050] The concatenation of the swin-transformer layer and the convolutional block layer can be expressed as:

[0051] Y=Cat(3D SwinT(X),3D Cov(X))

[0052] Among them, Cat() is the concatenation operation along the channel dimension; 3D SwinT(X) is the output of the swin-transformer layer, 3D Cov(X) is the output of the convolution block layer, and Y is the output of the 3D encoding block.

[0053] In this embodiment, by dividing the input features into two parts with the same number of channels and inputting them into the Swin-Transformer layer and the convolutional block layer respectively, learning of different scales and feature types is achieved.

[0054] The Swin-Transformer layer is good at capturing global features, and the convolutional block layer is suitable for extracting local detail features. By splicing the outputs of the Swin-Transformer layer and the convolutional block layer along the channel dimension, the features extracted by the two can be effectively fused to form a more comprehensive representation, retaining the complementary information of the two features and providing richer input for subsequent networks. The upsampling layer gradually enlarges the size of the feature map, restores the details layer by layer, and then the attention weighted feature fusion layer can weight the features at different positions, so that the decoder can focus more on the coronary artery area, thereby improving the separation of the coronary artery from the background, reducing noise interference, and generating clearer and more accurate segmentation results.

[0055] In an optional embodiment, the decoding block includes an upsampling layer and an attention weighted feature fusion layer connected in sequence.

[0056] like Figure 4As shown, this is the structural diagram of the attention weighted feature fusion layer.

[0057] In this embodiment, the spatial resolution of the image is gradually restored through the upsampling layer, so that the size of the feature map is closer to the original image, thereby providing basic support for restoring the details of the coronary arteries; secondly, the weights of different feature channels are adaptively adjusted through the attention weighted feature fusion layer to highlight the features of the coronary artery area, suppress the interference of background noise, and make the segmentation result more accurate.

[0058] Further optionally, the spatial attention layer obtains features through global average pooling and global maximum pooling, and performs fusion weighting through convolution; the channel attention layer obtains features through global average pooling, and further adjusts weights through continuous point convolution.

[0059] Specifically, the attention weighted feature fusion module includes spatial attention and channel attention. The spatial attention uses global average pooling and global maximum pooling respectively, and then uses 7×7×7 convolution to obtain W s ; Channel attention is achieved by global average pooling and continuous point convolution to obtain W C According to the broadcast mechanism, W is added by simple addition operation. c and W s Fuse together to get W cs ∈R C×H×W×D The summation result is applied to the Sigmoid function to obtain the attention coefficient α∈[0,1], and then the weighted summation method is used to fuse the output; its expression is as follows:

[0060]

[0061] W CS =W C +W S

[0062] Y=X1×(1-α)+X2×α

[0063] in, and Respectively represent the global maximum pooling along the spatial dimension, the global average pooling along the spatial dimension, and the global average pooling along the channel dimension; Conv k represents a convolution with a kernel size of k×k×k; R represents the Relu activation function; Cat represents the concatenation operation in the channel dimension; Y∈R C×H×W×D The fused features.

[0064] In this embodiment, through the combination of spatial attention and channel attention, the module can simultaneously focus on the spatial distribution of the feature map and the features of different channels. Based on the dual attention mechanism, the model can accurately capture the spatial position of the coronary arteries and the correlation between the channels, thereby more effectively separating the target area from the background. Among them, spatial attention extracts important features at different positions through a combination of global average pooling and maximum pooling, and then uses 7×7×7 convolution for fusion, making the spatial features richer and more concentrated. Channel attention strengthens the weights of key channels through the processing of global average pooling and continuous point convolution, which helps the model focus more on features that are beneficial to segmentation.

[0065] In an optional embodiment, the step of integrating the features through the multi-scale feature interaction module includes:

[0066] The features are divided into several different heads along the channel dimension, and convolutions with different dilation rates are used on the heads to extract multi-scale features of the coronary artery;

[0067] The multi-scale features of the heads are fused through point convolution to obtain integrated features.

[0068] like Figure 5 As shown in Figure 1, it is the structural diagram of the multi-scale feature interaction module.

[0069] Specifically, the multi-scale feature interaction module applies convolutions with different dilation rates to the input feature map in parallel. C×H×W×D Divided into four different heads along the channel dimension, generating Four 3×3×3 convolutions with different dilation rates are used to extract the multi-scale features of the coronary artery. X i ' is divided into separate channels, and for each head, we get The individual channels of each head are cross-reorganized into Then, point convolution is used to fuse the information between groups and between groups, and the output Y = ∈ R C ×H×W×D ; Its expression is as follows:

[0070] X i =DConv(X i )

[0071]

[0072] Y=R(B(Conv1(Cat(h1,h2,...,h j ))))

[0073] Where X i′ represents the features obtained by applying dilated convolution to the i-th head, DConv represents depthwise separable dilated convolution, and the dilation rates used are i=1, 3, 5, 7 respectively;

[0074] Where Conv1 represents a point convolution with a convolution kernel size of 1×1×1. represents the jth channel of the i-th head, h j represents the jth group of features R and B represent the Relu activation function and normalization function respectively.

[0075] In this embodiment, by dividing the features into multiple heads along the channel and using convolutions with different expansion rates, the features of the coronary arteries can be extracted at different scales, which helps the model capture subtle local information and global structural information at the same time; 3×3×3 convolutions with different expansion rates can enable the model to better adapt to the complexity of the coronary artery structure, effectively cope with the morphological changes of the coronary arteries in the image, and enhance the segmentation model's ability to identify vascular branches and intersections. Secondly, by fusing the multi-scale features of each head through point convolution, information can be fully exchanged between and within groups, thereby achieving deeper feature integration; point convolution also ensures the efficient flow of information between scales, making the final output features more representative and discriminative.

[0076] In an optional embodiment, the method further includes: acquiring a segmented three-dimensional coronary artery image, inputting the image into the neural network model and training based on back propagation until the neural network model converges; wherein the back propagation is based on dice loss and cross entropy loss; the loss function of the neural network model is as follows:

[0077]

[0078] Where N is the number of voxels, m is the number of output categories, and y ij represents the class prediction output, Represents the class-encoded target vector.

[0079] Further optionally, a region of interest is extracted from the divided three-dimensional coronary artery image, the CT value of the image is limited to a preset range, and the image is randomly cropped into n three-dimensional blocks of size a×a×a and input into the neural network model for training.

[0080] As an exemplary illustration, the preset range of CT value is [-260, 730] HU.

[0081] In this embodiment, the neural network is trained by segmenting the three-dimensional coronary artery image, and the loss function is used to calculate the loss between the segmented image and the coronary artery label. By combining the Dice loss and the cross entropy loss, the class imbalance problem can be effectively handled. Among them, the Dice loss focuses on the degree of regional overlap and has a stronger sensitivity to the segmentation of coronary arteries in small areas, which helps to improve the accuracy of the model for small coronary artery structures; the cross entropy loss can provide accurate gradient information for the classification of each pixel, so that the model converges faster.

[0082] In an optional embodiment, the method further includes: performing a 3D morphological closing operation on the segmentation result, calculating the volume of each connected domain, and removing areas whose volume is smaller than a preset threshold.

[0083] In this embodiment, the segmentation results are further optimized. Through the 3D morphological closing operation, small gaps and small breaks in the segmentation results can be filled, making the segmentation results of the coronary arteries more continuous and smooth, and improving the integrity of the segmentation boundaries; removing small areas with a volume smaller than a preset threshold can effectively remove noise and irrelevant artifacts, ensuring that the segmentation results are concentrated on the main structure of the coronary arteries, and improving the accuracy and clarity of the segmentation results.

[0084] Example 2

[0085] This embodiment proposes a 3D coronary artery segmentation system based on CNN and Transformer, and applies the 3D coronary artery segmentation method based on CNN and Transformer proposed in Example 1. Figure 6 , which is an architecture diagram of the CNN and Transformer-based three-dimensional coronary artery segmentation system of this embodiment.

[0086] This embodiment proposes a 3D coronary artery segmentation system based on CNN and Transformer, the system comprising:

[0087] Model building unit: constructing a neural network model including an encoder and a decoder based on CNN and Transformer, wherein the bottleneck layer in the neural network model includes a multi-scale feature interaction module;

[0088] Image segmentation unit: inputting the three-dimensional coronary artery image to be segmented into the neural network model; wherein, in the encoder, extracting the features of the three-dimensional coronary artery image to be segmented, integrating the features through the multi-scale feature interaction module at the bottleneck layer; in the decoder, weighted fusion of the integrated features and the sampling features output by each layer of the decoder is performed and the segmentation result is output.

[0089] It can be understood that the system of this embodiment corresponds to the method of the above-mentioned embodiment 1, and the options in the above-mentioned embodiment 1 are also applicable to this embodiment, so they will not be described repeatedly here.

[0090] Example 3

[0091] This embodiment uses the CNN and Transformer-based 3D coronary artery segmentation method proposed in Embodiment 1 to segment the 3D coronary artery image.

[0092] like Figure 7 The actual segmentation result of a 3D coronary artery image is shown in Figure 8 The figure shows the result of segmenting a three-dimensional coronary artery image using a three-dimensional coronary artery segmentation method based on CNN and Transformer proposed in Example 1.

[0093] Depend on Figure 7 and Figure 8 It can be seen that the segmentation results of the 3D coronary artery segmentation method based on CNN and Transformer in this application capture the main trunk and main branches of the blood vessels, maintain good continuity in the connection and extension of the branches, and have no fragments or isolated points, effectively solving the overall continuity problem of the vascular structure and reducing the irregular edges caused by fragmentation. Figure 8 The boundary of the segmentation result is relatively smooth, without obvious jagged edges or abrupt breakpoints, which helps to improve the visual consistency of the segmentation and conforms to the natural smooth shape of the coronary artery.

[0094] The Dice, Recall, Precision, and HD95 indicators are used to evaluate the similarity between the segmentation results of the classic 3D Unet, Vnet, Att-Unet, SegresNet, Unetr, Swin-unetr, 3D UX-Net, and the model of this application and the real label. The Dice coefficient is a commonly used overlap metric used to measure the degree of overlap between the predicted segmentation area and the real segmentation area. The value range of the Dice coefficient is 0 to 1. The larger the value, the higher the overlap between the segmentation result and the real label. The Recall coefficient measures the ability of the segmentation model to identify all real positive samples. The higher the Recall coefficient, the fewer real targets the model misses during segmentation. The Precision coefficient measures the proportion of positive samples predicted by the segmentation model that are truly positive samples. The higher the Precision coefficient, the lower the probability of the model making mistakes when predicting positive samples. The HD95 coefficient is used to measure the distance between the predicted boundary and the real label boundary. The smaller the HD95 coefficient, the closer the boundary of the segmentation result is to the real boundary.

[0095] Table 1 shows the index data of the segmentation results obtained by all the above models. It can be seen that the segmentation results of the model of this application are greater than those of other models in terms of dice coefficient, smaller than those of other models in terms of HD95 coefficient, and similar to those of other models in terms of recall coefficient and precision coefficient; this shows that compared with the prior art, this application has a higher level of segmentation and the accuracy is comparable to that of the prior art.

[0096] Table 1 Index data of segmentation results under different models

[0097]

[0098] Example 4

[0099] This embodiment proposes a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, wherein when the computer-readable instructions are executed by the processor, the processor executes the steps of a three-dimensional coronary artery segmentation method based on CNN and Transformer proposed in Example 1.

[0100] The terms used in the drawings are for illustrative purposes only and should not be construed as limiting this patent;

[0101] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A 3D coronary artery segmentation method based on CNN and Transformer, characterized in that: The following steps are involved: A neural network model including an encoder and a decoder is constructed based on CNN and Transformer, wherein the bottleneck layer in the neural network model includes a multi-scale feature interaction module; The three-dimensional coronary artery image to be segmented is input into the neural network model; wherein, the features of the three-dimensional coronary artery image to be segmented are extracted in the encoder, and the features are integrated through the multi-scale feature interaction module in the bottleneck layer; and in the decoder, the integrated features and the sampling features output by each layer of the decoder are weightedly fused and the segmentation result is output.

2. The three-dimensional coronary artery segmentation method based on CNN and Transformer according to claim 1, characterized in that: The encoder and decoder adopt a multi-layer symmetrical structure. The encoder includes a plurality of 3D encoding blocks connected in sequence; the decoder includes a plurality of decoding blocks connected in sequence.

3. The three-dimensional coronary artery segmentation method based on CNN and Transformer according to claim 2, characterized in that: The 3D coding block includes a swin-transformer layer and a convolution block layer; the features in the 3D coding block are divided into two features with the same number of channels and input into the swin-transformer layer and the convolution block layer respectively, and the outputs of the swin-transformer layer and the convolution block layer are spliced ​​as the output of the 3D coding block.

4. The three-dimensional coronary artery segmentation method based on CNN and Transformer according to claim 2, characterized in that: The decoding block includes an upsampling layer and an attention-weighted feature fusion layer connected in sequence.

5. The three-dimensional coronary artery segmentation method based on CNN and Transformer according to claim 4, characterized in that: The attention weighted feature fusion layer includes a spatial attention layer and a channel attention layer; the spatial attention layer obtains features through global average pooling and global maximum pooling, and performs fusion weighting through convolution; the channel attention layer obtains features through global average pooling, and further adjusts the weights through continuous point convolution.

6. The three-dimensional coronary artery segmentation method based on CNN and Transformer according to claim 1, characterized in that: The step of integrating the features through the multi-scale feature interaction module includes: The features are divided into several different heads along the channel dimension, and convolutions with different dilation rates are used on the heads to extract multi-scale features of the coronary artery; The multi-scale features of the plurality of heads are fused through point convolution to obtain integrated features.

7. The three-dimensional coronary artery segmentation method based on CNN and Transformer according to any one of claims 1 to 6, characterized in that: The method further includes: obtaining a segmented three-dimensional coronary artery image, inputting the image into the neural network model and training based on back propagation until the neural network model converges; wherein the back propagation is based on dice loss and cross entropy loss; the loss function of the neural network model is as follows: Where N is the number of voxels, m is the number of output categories, and y ij represents the class prediction output, Represents the class-encoded target vector.

8. The three-dimensional coronary artery segmentation method based on CNN and Transformer according to any one of claims 1 to 6, characterized in that: The method further comprises: performing a 3D morphological closing operation on the segmentation result, calculating the volume of each connected domain, and removing areas whose volume is smaller than a preset threshold.

9. A 3D coronary artery segmentation system based on CNN and Transformer, applied to the 3D coronary artery segmentation method based on CNN and Transformer according to any one of claims 1 to 8, characterized in that: The system comprises: Model building unit: constructing a neural network model including an encoder and a decoder based on CNN and Transformer, wherein the bottleneck layer in the neural network model includes a multi-scale feature interaction module; Image segmentation unit: inputting the three-dimensional coronary artery image to be segmented into the neural network model; wherein, in the encoder, extracting the features of the three-dimensional coronary artery image to be segmented, integrating the features through the multi-scale feature interaction module at the bottleneck layer; in the decoder, weighted fusion of the integrated features and the sampling features output by each layer of the decoder is performed and the segmentation result is output.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the three-dimensional coronary artery segmentation method based on CNN and Transformer is implemented as described in any one of claims 1-8.

Images