Medical image segmentation method and system based on space-contour attention network

Abstract

The invention discloses a medical image segmentation method and system based on a space-contour attention network, and the method comprises the steps: designing an image segmentation model based on a U-shaped network, adding an edge contour extraction module in an encoder, and replacing a jump connection module in the original U-shaped network with a space-contour attention module. The edge contour extraction module is used for modeling the edge information of the blood vessel and extracting a remarkable edge feature, and the edge feature is used as a key attention area in a network segmentation process; the space-contour attention module enables the network to pay attention to the space information and edge contour information of the blood vessel, the isolation of the blood vessel and the background, and the inhibition of the background features when the network carries out the segmentation task, thereby improving the representation capability of the network, reducing the front and back disconnection phenomena during the segmentation of the blood vessel, and enhancing the segmentation effect of the whole blood vessel. Experiments show that compared with an existing method, the coronary artery angiography image segmentation method can effectively improve the performance of coronary artery angiography image segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, in particular to a medical image segmentation method and system based on a spatial-contour attention network. Background Art

[0002] With the increasing emphasis on the diagnosis and prevention of cardiovascular diseases, digital subtraction angiography technology, as an accurate examination method, has received more and more attention. Automatically extracting a clear and complete vascular tree structure from coronary angiography images plays a significant role in the study of coronary artery diseases: highly accurate vascular segmentation results are helpful for calculating data such as the stenosis length and blood flow velocity of coronary arteries in scientific research, and at the same time are helpful for subsequent three-dimensional modeling and analysis work on the affected area; in clinical applications, it is helpful for assisting doctors to diagnose the affected area of patients and formulate the next diagnosis and treatment plan. Therefore, the segmentation of coronary arteries is of great significance for the precise quantification and diagnosis of coronary artery diseases.

[0003] Currently, the segmentation of coronary arteries still relies on the subjective annotation results of experts. The annotation process consumes a large amount of manpower and time, and there are individual differences in the annotation results, which are likely to affect the diagnosis of doctors. In recent years, the rapid development of big data and artificial intelligence technologies in China has laid a solid foundation for the technology of automatically extracting coronary arteries by computers. Compared with the manual annotation of medical experts, the computer automatic extraction technology has incomparable advantages: on the one hand, the automatic extraction technology is not affected by human subjective factors; on the other hand, automatic segmentation can reduce time and economic costs.

[0004] Traditional segmentation methods usually divide an image into regions according to features such as the color, texture, and shape of the image itself, highlight the differences between different regions to separate the background, and finally obtain a vascular tree with a single color. Compared with traditional methods, the deep learning method approximates the label through continuous training and iteration, which ensures the accuracy of the segmentation result. As the training set expands, the training accuracy will continue to improve, and it has strong robustness for images with unified characteristics. Compared with the disadvantage of long time consumption of traditional methods, the deep learning-based method has the characteristics of fast calculation speed and strong adaptability, and can process images in batches.

[0005] In deep learning, Attention-UNet can enhance the model's attention to important regions to a certain extent by introducing an attention mechanism. Especially when dealing with complex vascular crossing regions, it can reduce the interference of background noise. However, the limitation of Attention-UNet is that its attention mechanism is mainly based on the weighting of global information and lacks sufficient attention to local details (such as the ends of blood vessels or tiny branches). Especially for relatively complex blood vessel ends and small branches, Attention-UNet often fails to achieve satisfactory segmentation accuracy, and it is prone to missed detection or blurred edges.

[0006] SA-UNet (Spatial Attention U-Net) further enhances the utilization of spatial information on this basis and mines the spatial features in the vascular region through a spatial attention mechanism. SA-UNet shows good segmentation results when dealing with most vascular regions, especially in the recognition of the boundaries and contours of blood vessels. However, the spatial attention mechanism of SA-UNet mainly relies on the spatial information of shallow features and often fails to deeply capture the high-level semantic information of blood vessels. At the same time, SA-UNet is still insufficient in the segmentation of blood vessel ends and smaller branches. Especially in some complex regions or images with low contrast, it is prone to cause breaks or incoherence in the segmentation results. Summary of the Invention

[0007] Objective of the Invention: The objective of the present invention is to provide a medical image segmentation method and system based on a spatial-contour attention network. A spatial-contour attention module is designed according to the characteristics of coronary angiography blood vessels to improve the semantic segmentation effect of the original U-shaped network or its improvement on coronary angiography images.

[0008] Technical Solution: To achieve the above objective of the invention, the present invention adopts the following technical solutions:

[0009] In the first aspect, the present invention provides a medical image segmentation method based on a spatial-contour attention network, including the following steps:

[0010] Extract the edge contours of the segmentation labels of coronary angiography images to generate edge contour labels;

[0011] Construct a spatial-contour attention network based on the U-shaped network framework. Add an edge contour extraction module to the encoder module of the U-shaped network and replace the skip connection module of the original U-shaped network with a spatial-contour attention module; the edge contour extraction module extracts vascular edge contour features from the features of different layers in the main network, and the edge contour features and the features of different layers in the main network are added to the spatial-contour attention module to generate vascular adaptive weights;

[0012] The spatial-profile attention network is trained based on a dataset including segmentation labels and edge contour labels to obtain a converged network model. The trained network model is input with test coronary angiography images to obtain segmentation results.

[0013] Preferably, the edge contour extraction module is used to perform scale assimilation on the shallow and deep feature layers, fuse the feature layers after assimilating the scales through convolution, and generate edge contour features unique to the current segmentation instance under the supervision of contour labels; the edge contour features are used to fuse with feature layers at different levels in the encoder in the spatial-profile attention module.

[0014] Preferably, the encoder of the U-shaped network contains four layers. The edge contour extraction module selects the first-layer feature layer and the third-layer feature layer in the encoder, performs upsampling and activation operations on the third-layer feature layer so that the size of the third-layer feature layer becomes twice the original and the number of channels becomes half, and repeats this operation until the size of the third-layer feature layer is the same as that of the first-layer feature layer; the two feature layers are concatenated and convolved to be fused into a new feature layer with edge contour features. After two convolution and ReLU activation operations on this new feature layer, convolution and Sigmoid activation are performed to form a mapping graph of edge contour features.

[0015] Preferably, the spatial-profile attention module consists of spatial attention and edge contour attention; the spatial attention fuses convolution after performing max pooling and average pooling on the feature layers in the encoder respectively in the channel dimension to obtain a spatial attention weight map; the edge attention weight map is the edge contour features generated by the edge contour extraction module, and the same pooling operation as that of the feature layers in the encoder is adopted, and its size is consistent with that of the feature layers; the spatial-profile attention module is connected to the encoder and the decoder, with the input being different feature layers in the encoder and the edge contour features generated by the edge contour extraction module, and the output being feature layers with spatial-profile feature information; the spatial attention weight map and the edge attention weight map are weighted and summed and then normalized to obtain a spatial-profile attention weight map.

[0016] Preferably, the training steps of the spatial-profile attention network mainly include two paths. The main path is the training vascular segmentation path, and the auxiliary path is the training edge segmentation path. And the training of the main path and the auxiliary path is in the same training cycle. First, the training of the auxiliary path is completed to obtain edge contour features to assist the training of the main path;

[0017] Auxiliary path training process: The input image undergoes multiple convolutional, ReLU activation, and pooling operations in the encoder to obtain a deep feature layer. The feature layer is then upsampled to restore the spatial resolution of the image and convolved with the shallow feature layer. After passing through the activation function, the predicted edge contour label after training is obtained, which is compared with the input edge contour label. The loss function used is binary cross-entropy loss;

[0018] Main path training process: The input image extracts low-level local features and high-level global features through the encoder. The edge contour features generated in the auxiliary path are pooled together with the feature layer to keep the same size. In the decoder stage, the spatial resolution of the image is gradually restored through upsampling. At the same time, the spatial-contour attention module combines the feature layer in the encoder with the edge contour features and convolves and fuses them with the feature layer in the decoder; finally, the predicted segmentation label after training is obtained through the activation function, which is compared with the input label. The loss function used is binary cross-entropy loss.

[0019] In a second aspect, the present invention provides a medical image segmentation system based on a spatial-contour attention network, including:

[0020] A preprocessing module for extracting edge contours from the segmentation label of the coronary angiography image to generate an edge contour label;

[0021] A network model construction module, where the user constructs a spatial-contour attention network based on the U-shaped network framework, adds an edge contour extraction module to the encoder module of the U-shaped network, and replaces the skip connection module of the original U-shaped network with a spatial-contour attention module; the edge contour extraction module generates vascular edge contour features by extracting features from different layers in the main network, and the edge contour features and features from different layers in the main network are added to the spatial-contour attention module to generate vascular adaptive weights;

[0022] A network training and testing module for training the spatial-contour attention network based on a dataset including segmentation labels and edge contour labels to obtain a converged network model, and inputting the trained network model with a test coronary angiography image to obtain a segmentation result.

[0023] In a third aspect, the present invention provides a computer system, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, the steps of the medical image segmentation method based on the spatial-contour attention network are implemented.

[0024] In a fourth aspect, the present invention provides a computer program product, including a computer program. When the computer program is executed by the processor, the steps of the medical image segmentation method based on the spatial-contour attention network are implemented.

[0025] Beneficial effects: A medical image segmentation method based on a spatial-contour attention network provided by the present invention designs an improved U-shaped network based on the basic structure of the U-shaped network as a segmentation model. An edge contour extraction module is added to the encoder module, and the skip connection module of the original U-shaped network is replaced with a spatial-contour attention module. Among them, the edge contour extraction module is to model the edge information of blood vessels and extract significant edge features, which will be used as the key attention areas in the network segmentation process and provide conditions for the spatial-contour attention module; the spatial-contour attention module is to enable the network to focus on the spatial information and edge contour information of blood vessels during the segmentation task, pay attention to the isolation of blood vessels from the background, reduce the phenomenon of front-back disconnection during blood vessel segmentation, and enhance the overall segmentation effect of blood vessels. Compared with the prior art, the present invention has the following advantages:

[0026] 1. A spatial-contour attention mechanism is proposed for the fracture and discontinuity phenomena generated during the blood vessel segmentation of coronary angiography images. This mechanism enhances the attention of the network to the edge contours of blood vessels during blood vessel segmentation, increases the weight of blood vessel edge features in the network, and suppresses background features, which is beneficial to the integrity and continuity of blood vessel segmentation in coronary angiography images.

[0027] 2. By adding an edge contour extraction module to the U-shaped network, the network spontaneously generates edge contour features, and the edge contour features are input into the spatial-contour attention module to improve the segmentation performance of coronary angiography images. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is the implementation flowchart of the present invention;

[0029] Figure 2 is the schematic diagram of the network structure of the present invention;

[0030] Figure 3 is the schematic diagram of the structure of the spatial-contour attention module;

[0031] Figure 4 is the result graph of the comparison between the present invention and the existing method on the dataset. DETAILED DESCRIPTION OF THE INVENTION

[0032] The technical solutions and effects of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0033] A medical image segmentation method based on a spatial-contour attention network disclosed in an embodiment of the present invention has an implementation flowchart as shown in Figure 1 and specifically includes the following steps:

[0034] Step 1: Image preprocessing: Crop the coronary angiography images and their corresponding segmentation labels to remove the redundant background, extract the edge contours of the segmentation labels to generate new edge contour labels, which are used in the network to train the edge contour features.

[0035] Step 2: Construct a spatial-contour attention network: First, construct a U-shaped network framework. Based on the U-shaped network, add an edge contour extraction module and a spatial-contour attention module. The edge contour extraction module generates vascular edge contour features by extracting features from different layers in the main network. The edge contour features and the features from different layers in the main network are added to the spatial-contour attention module to generate more effective vascular adaptive weights.

[0036] Specifically, in this embodiment, the spatial-contour attention network includes an encoding network, a decoding network, an edge contour extraction module, and a spatial-contour attention module. The encoding network extracts features from the image. The edge contour extraction module decodes the features in the encoder to generate an edge contour feature map. The edge contour feature map and the features of different scales in the encoder are combined in the spatial-contour attention module to generate new features of different scales that focus on the vascular edge contours. The features of different scales are sent into the decoding network and are organically combined with the feature maps in the decoding network in a convolutional manner to fully extract the edge spatial information of the blood vessels.

[0037] As Figure 2 shown, in this embodiment, the encoding network contains four layers, each layer is composed of convolution, and is generated through pooling operations between layers; the decoding network contains four layers, each layer corresponds to the encoding network, and is generated through deconvolution operations between layers; the encoding network and the decoding network are connected through the spatial-contour attention module, and the spatial-contour attention module is also divided into four layers, which connect the corresponding layers of the encoding network and the decoding network respectively.

[0038] The edge contour extraction module contains two layers, and is generated through deconvolution operations between layers and is connected to the first layer and the third layer of the encoding network, and the generated edge contour feature map is connected to the spatial-contour attention module.

[0039] The edge contour extraction module mainly includes scale assimilation of the shallow and deep feature layers, fusing the feature layers after assimilating the scales through convolution, and generating edge contour features unique to the current segmentation instance through contour label supervision. The edge contour features are used to fuse with the feature layers of different levels in the encoder in the spatial-contour attention module, assisting the network to focus on the edge contours during segmentation, and strengthening the constraint of the edge contours on the segmentation result.

[0040] Specifically, in the encoder, a feature layer with 64 channels and a feature layer with 256 channels in the deep layer are selected. The deep feature layer is upsampled and subjected to ReLU activation operations so that its size becomes twice the original and the number of channels becomes half. This operation is repeated until the size of the feature layer is the same as that of the shallow feature layer. The two feature layers are concatenated and convolved to be fused into a new feature layer with rich edge contour features. After two 3*3 convolution ReLU activation operations on this feature layer, a 1*1 convolution and Sigmoid activation are performed to form a mapping graph of edge contour features.

[0041] This process is represented by the following formula:

[0042]

[0043] where C (1) represents the shallow feature layer, represents the deep feature layer, represents the convolution operation of the deep feature, θ is the convolution parameter, φ() represents the ReLU activation function, Up(*; C (1) ) represents converting * into a feature layer with the same size as C (1) through upsampling.

[0044] As Figure 3 shown, the spatial-contour attention module consists of spatial attention and edge contour attention. The spatial attention is obtained by performing max pooling and average pooling on the feature layer in the encoder along the channel dimension respectively and then fusing the convolution to obtain the spatial attention weight map; the edge attention weight map is the edge contour feature generated by the edge contour extraction module. The edge contour attention weight map performs the same pooling operation as the feature layer in the encoder, and its size is consistent with the feature layer, and the number of channels is 1;

[0045] The spatial-contour attention module is connected to the encoder and the decoder. The input is different feature layers in the encoder and the edge contour features generated by the edge contour extraction module, and the output is a feature layer with spatial-contour feature information;

[0046] The spatial attention weight map and the edge attention weight map are weighted and summed and then normalized to obtain the spatial-contour attention weight map. This weight map contains the spatial information of the image and the important features of the edge contour. The weight map is fused with the feature layer generated in the encoder to obtain a feature layer with edge contours and spatial features, and these feature layers are finally fed into the decoder.

[0047] The spatial-contour attention module is as shown in the formula:

[0048] F ESA = F·σ(f 7×7([MaxPool(F); AvgPool(F)]) + αF E )

[0049] where F ESA is the generated spatial - contour feature layer, F is the feature layer in the original encoder, and f 7×7 is the 7×7 convolution operation, MaxPool(F) and AvgPool(F) are the max - pooling and average - pooling operations, and F E is the edge - contour feature generated in the edge - contour extraction module, α is the weight value, and σ is the sigmoid activation function.

[0050] Step 3: Train the network: Divide the data processed in Step 1 into a training set, a test set, and a validation set, and use the data to train the network model to obtain a converged network model; during the training process, the input image data to the network generates not only the segmentation image but also the edge - contour image of blood vessels compared with the original U - shaped network, and the generated edge - contour feature image will be fed into the spatial - contour attention module by the network to assist in the generation of the final segmentation image.

[0051] The training steps of the spatial - contour attention network mainly include two paths. The main path is to train the blood - vessel segmentation path, and the auxiliary path is to train the edge - segmentation path. And the training of the main path and the auxiliary path is in the same training cycle. First, complete the training of the auxiliary path to obtain the edge - contour feature to assist the training of the main path. The loss function uses Binary Cross - Entropy Loss.

[0052] Training process of the auxiliary path: The input image and the corresponding edge - contour label are input into the network together and trained in the edge - segmentation path. The input image undergoes three convolution, ReLU activation, and pooling operations in the encoder to obtain a deep - layer feature layer. The spatial resolution of the image is restored by performing two up - sampling operations on this feature layer and convolving it with the shallow - layer feature layer, and then passing through the sigmoid activation function to obtain the predicted edge - contour label after training. Compare it with the input label, and the loss function used is Binary Cross - Entropy Loss:

[0053]

[0054] where y edge,i is the edge - contour label, is the probability value of the edge region output by the network, and N is the number of samples.

[0055] Main path training process: The input image and the corresponding segmentation label are input into the network together. The encoder in the network is responsible for extracting low-level local features and high-level global features from the input image. This stage includes multiple convolution operations and pooling operations. The edge contour features generated in the auxiliary path are pooled along with the feature layers to keep the same size. In the decoder stage, the network gradually restores the spatial resolution of the image through upsampling. At the same time, the spatial-contour attention module in the decoder combines the feature layers in the encoder with the edge contour features and convolves and fuses them with the feature layers in the decoder to improve the network's attention to the blood vessel contour. Finally, the trained predicted segmentation label is obtained through the sigmoid activation function and compared with the input label. The loss function used is binary cross-entropy loss:

[0056]

[0057] where y i is the segmentation label, is the probability value of the blood vessel area output by the network.

[0058] Step 4: Coronary angiography image segmentation: The coronary angiography image to be processed is input into the network for fully automatic coronary artery segmentation.

[0059] The performance of the network model is evaluated by evaluation metrics Precision, Recall, F1-score, Intersection over Union (IOU), and Dice-Score:

[0060] Precision represents the proportion of samples predicted as positive by the model that are predicted correctly (true label is positive), representing the accuracy of the model during segmentation.

[0061] Recall represents what percentage of all samples with true labels as positive are predicted, that is, for the blood vessels in the true labels, how many of the model's segmentation results are judged as blood vessel parts.

[0062] F1-score is the harmonic mean of precision and recall, comprehensively considering the influence of both, providing a balanced evaluation criterion. It is applicable to situations where there is a trade-off between precision and recall.

[0063] Intersection over Union (IOU) is the intersection-over-union of the predicted samples and the actual samples. Dice-Score is an index used to evaluate the similarity between the predicted segmentation map and the true label in an image segmentation task. It is defined as twice the intersection of the two sets divided by the sum of the two sets.

[0064] Dice-Score is a metric used to evaluate the similarity between the predicted segmentation map and the ground truth label in image segmentation tasks. It is defined as twice the intersection of two sets divided by the sum of the two sets.

[0065] To objectively demonstrate the performance of the method proposed in the present invention in coronary angiography image segmentation, the network was compared and evaluated with other representative medical segmentation models, such as the U-Net-DB, Attention-UNet, SA-UNet network models, etc. The experimental data of the network models are shown in Table 1:

[0066] Table 1 Comparison of evaluation metrics for the segmentation results of the algorithms

[0067] Model IOU Dice Precision Recall F1 U-Net 0.810 0.686 0.914 0.877 0.895 U-Net-DB 0.827 0.697 0.921 0.890 0.905 Attention-UNet 0.835 0.704 0.926 0.894 0.909 SA-UNet 0.817 0.693 0.917 0.881 0.898 This method 0.862 0.718 0.936 0.915 0.925

[0068] As shown in Table 1, Figure 4 as described, when compared and tested with other methods on the JMA dataset, the method of the present invention achieved the best performance.

[0069] The embodiment of the present invention also discloses a medical image segmentation system based on a spatial-contour attention network, including: a preprocessing module for extracting the edge contours of the segmentation labels of coronary angiography images to generate edge contour labels; a network model construction module for constructing a spatial-contour attention network based on the U-shaped network framework, adding an edge contour extraction module to the encoder module of the U-shaped network, and replacing the skip connection module of the original U-shaped network with a spatial-contour attention module; the edge contour extraction module generates vascular edge contour features by extracting features from different layers in the main network, and the edge contour features and features from different layers in the main network are input into the spatial-contour attention module to generate vascular adaptive weights; a network training and testing module for training the spatial-contour attention network based on a dataset including segmentation labels and edge contour labels to obtain a converged network model, and inputting the trained network model with test coronary angiography images to obtain segmentation results. The specific implementation details of each module can be referred to the above method embodiment and will not be elaborated here.

[0070] The embodiment of the present invention also discloses a computer system, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the medical image segmentation method based on the spatial-contour attention network.

[0071] The embodiment of the present invention also discloses a computer program product, including a computer program. When the computer program is executed by the processor, it implements the steps of the medical image segmentation method based on the spatial-contour attention network.

[0072] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or a controller of a general-purpose computer, a special-purpose computer, or other programmable data processing devices, so that when the program codes are executed by the processor or the controller, the steps of the method of the present invention are implemented. The program codes can be executed entirely on the machine, partially on the machine, partially on the machine as an independent software package and partially on a remote machine, or entirely on a remote machine or a server. Where the present invention is not described in detail, it is well-known technology to those skilled in the art.

Claims

1. A medical image segmentation method based on a spatial-contour attention network, characterized in that: The following steps are involved: Extract edge contours from segmentation labels of coronary angiography images to generate edge contour labels; A space-contour attention network is constructed based on a U-shaped network framework, an edge contour extraction module is added to the encoder module of the U-shaped network, and the skip connection module of the original U-shaped network is replaced by the space-contour attention module; the edge contour extraction module generates blood vessel edge contour features by extracting features from different layers in the main network, and the edge contour features and features from different layers in the main network are added to the space-contour attention module to generate blood vessel adaptive weights; The spatial-contour attention network is trained based on a data set including segmentation labels and edge contour labels to obtain a converged network model. The trained network model is input with a test coronary angiography image to obtain the segmentation result.

2. The medical image segmentation method based on the space-contour attention network according to claim 1, characterized in that: The edge contour extraction module is used to scale the shallow and deep feature layers, fuse the assimilated feature layers through convolution, and generate edge contour features unique to the current segmentation instance through contour label supervision; The edge contour features are used to fuse with the feature layers of different levels of the encoder in the network in the spatial-contour attention module.

3. The medical image segmentation method based on space-contour attention network according to claim 2, characterized in that: The encoder of the U-shaped network contains four layers. The edge contour extraction module selects the first and third feature layers in the encoder, upsamples and activates the third feature layer so that the size of the third feature layer becomes twice as large as the original one and the number of channels becomes half. This operation is repeated until the size of the third feature layer is the same as that of the first feature layer. The two feature layers are concatenated and convolved to form a new feature layer with edge contour features. The new feature layer is convolved and ReLU activated twice, and then convolved and Sigmoid activated to form a mapping map of the edge contour features.

4. The medical image segmentation method based on space-contour attention network according to claim 1, characterized in that: The spatial-contour attention module consists of spatial attention and edge contour attention; the spatial attention obtains a spatial attention weight map by performing maximum pooling and average pooling on the feature layer in the encoder in the channel dimension and then fusing convolutions; the edge attention weight map is the edge contour feature generated by the edge contour extraction module, which adopts the same pooling operation as the feature layer in the encoder, and its size is consistent with the feature layer; the spatial-contour attention module is connected to the encoder and the decoder, and its input is the edge contour features generated by different feature layers in the encoder and the edge contour extraction module, and its output is a feature layer with spatial-contour feature information; the spatial attention weight map and the edge contour attention weight map are weighted summed and normalized to obtain the spatial-contour attention weight map.

5. The medical image segmentation method based on space-contour attention network according to claim 1, characterized in that: The training steps of the spatial-contour attention network mainly include two paths. The main path is the path for training blood vessel segmentation, and the auxiliary path is the path for training edge segmentation. The training of the main path and the auxiliary path are in the same training cycle. The training of the auxiliary path is completed first, and the edge contour features are obtained to assist the training of the main path. Auxiliary path training process: The input image undergoes multiple convolutions, ReLU activations, and pooling operations in the encoder to obtain a deep feature layer. The feature layer is upsampled to restore the spatial resolution of the image and convolved with the shallow feature layer through an activation function to obtain the edge contour label predicted after training, which is compared with the input edge contour label. The loss function used is binary cross entropy loss. Main path training process: The input image is extracted through the encoder to extract the bottom-level local features and high-level global features. The edge contour features generated in the auxiliary path are pooled together with the feature layer to keep the same size. In the decoder stage, the spatial resolution of the image is gradually restored through upsampling. At the same time, the spatial-contour attention module combines the feature layer in the encoder with the edge contour features and convolutionally fuses them with the feature layer in the decoder. Finally, the predicted segmentation label after training is obtained through the activation function and compared with the input label. The loss function used is the binary cross entropy loss.

6. A medical image segmentation system based on spatial-contour attention network, characterized in that: include: A preprocessing module is used to extract edge contours from segmentation labels of coronary angiography images to generate edge contour labels; Network model construction module, the user builds a space-contour attention network based on the U-shaped network framework, adds an edge contour extraction module to the encoder module of the U-shaped network, and replaces the jump connection module of the original U-shaped network with the space-contour attention module; the edge contour extraction module generates blood vessel edge contour features by extracting features from different layers in the main network, and the edge contour features and features from different layers in the main network are added to the space-contour attention module to generate blood vessel adaptive weights; The network training and testing module is used to train the space-contour attention network based on a data set including segmentation labels and edge contour labels to obtain a converged network model, and input a test coronary angiography image into the trained network model to obtain a segmentation result.

7. The medical image segmentation system based on space-contour attention network according to claim 6, characterized in that: The edge contour extraction module is used to scale the shallow and deep feature layers, fuse the assimilated feature layers through convolution, and generate edge contour features unique to the current segmentation instance through contour label supervision; The edge contour features are used to fuse with the feature layers of different levels of the encoder in the network in the spatial-contour attention module.

8. The medical image segmentation system based on space-contour attention network according to claim 6, characterized in that: The spatial-contour attention module consists of spatial attention and edge contour attention; the spatial attention obtains a spatial attention weight map by performing maximum pooling and average pooling on the feature layer in the encoder in the channel dimension and then fusing convolutions; the edge attention weight map is the edge contour feature generated by the edge contour extraction module, which adopts the same pooling operation as the feature layer in the encoder, and its size is consistent with the feature layer; the spatial-contour attention module is connected to the encoder and the decoder, and its input is the edge contour features generated by different feature layers in the encoder and the edge contour extraction module, and its output is a feature layer with spatial-contour feature information; the spatial attention weight map and the edge attention weight map are weighted summed and normalized to obtain the spatial-contour attention weight map.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of the medical image segmentation method based on the space-contour attention network according to any one of claims 1 to 5 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the medical image segmentation method based on the space-contour attention network according to any one of claims 1 to 5 are implemented.

Images