Coronary artery segmentation method combining MambaASPP and combined loss function

Abstract

The invention discloses a coronary artery segmentation method in combination with MambaASPP and a combined loss function. The method comprises the following steps: 1, through non-zero data cutting and resampling adjustment, computing resources are saved, and a model can better process images with different resolutions; and 2, on the basis of the U-Net model, a residual structure is added to each block in an encoder and a decoder, and then a MambaASPP module is added to a bottleneck layer, so that local boundary positioning and background and foreground contrast are enhanced, and small blood vessel segmentation accuracy is improved. And 3, in the U architecture of the model, adding a segmentation head to each layer to obtain a segmentation result, and supervising the module of each layer through a loss function to enable the model to learn detail information in advance. And 4, supervising model training: combining the cross entropy, Dice and cbDice according to a weight ratio of 3: 1: 2 to form a loss function, supervising from different angles, and improving the segmentation effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image segmentation, and specifically proposes a coronary artery segmentation method based on deep learning, which combines MambaASPP and a combined loss function, and ultimately improves the accuracy of small blood vessel segmentation, boundary segmentation and local connectivity recognition. Background Art

[0002] The incidence of cardiovascular disease has been increasing year by year, and it has become one of the most common causes of death in the world, seriously threatening human health. The number of people who die from cardiovascular disease and its complications each year has exceeded the total number of deaths from all cancers, among which coronary heart disease has a very high mortality rate. The main cause of coronary heart disease is coronary artery stenosis. The coronary artery is responsible for supplying blood to the heart, and its abnormality can be life-threatening. Assessing the degree of coronary artery stenosis is crucial in the diagnosis of coronary artery disease.

[0003] Currently, computed tomography angiography (CTA) is considered the best imaging method for the coronary arteries of the heart. It provides non-invasive, high-resolution 3D images and generates multiple heart slices. Manual analysis of the degree of stenosis in these slices is cumbersome and time-consuming. In addition, the slices also contain non-coronary artery structures, which may interfere with the judgment of stenosis. Therefore, automatic segmentation of coronary arteries in CTA can help experts more accurately assess cardiovascular conditions and improve early diagnosis and treatment effects. Most coronary artery segmentation methods still rely on semi-automatic methods, where experts guide the algorithm and correct errors, which is time-consuming and relies on expert experience. Therefore, there is an urgent need for effective automatic segmentation algorithms to assist in the diagnosis of cardiovascular diseases.

[0004] With the development of deep learning, semantic segmentation by learning the category of each pixel in CT images through deep neural networks has been widely used in coronary artery segmentation. In the field of medical image segmentation, CNN and Transforme have achieved a prominent position. CNN models such as U-Net and FCN can effectively extract local features and extract the structure of the region of interest to a large extent, and are widely used in medical image segmentation. Transformer was originally designed for natural language processing, but has been successfully applied to the field of image processing, such as Vision Transformer for image classification and recognition, and Swin Transformer, a general backbone network for various visual tasks, which is good at capturing global information. Due to the complementary characteristics of these two models, in many studies, the two are mixed and the images are processed together, such as UNETR, SwinUNETR, nnFormer. However, the time overhead of Transformer is quadratic to the size of the input image, and high-resolution medical image processing is expensive. Therefore, how to effectively enhance the ability to capture global information in CNN without making the overhead too large is a problem that needs to be solved. Recently, state-space sequence models (SSM), especially structured state-space sequence models (S4), have been regarded as effective modules for building efficient deep learning networks. Mamba further improves S4 through the selection mechanism and combines hardware perception technology to reduce time overhead. Mamba has been successfully applied to image processing, such as Vision Mamba and EfficientVMamba, and has achieved good results.

[0005] However, coronary artery segmentation faces some challenges: on the one hand, due to differences in anatomical structures, different branches of coronary arteries have different blood vessel diameters, and traditional loss functions such as Dice loss are more inclined to large blood vessels, resulting in insufficient sensitivity in the identification of small blood vessels. CT images usually have low contrast, which makes the boundary between coronary arteries and surrounding tissues blurred, increasing the difficulty of the segmentation task; in addition, artifact interference affects the judgment of local connectivity, which is very important for blood flow analysis and lesion detection. Summary of the invention

[0006] In order to solve the above-mentioned difficulties in coronary artery segmentation, the present invention provides a coronary artery segmentation method combining MambaASPP (Mamba AtrousSpatial Pyramid Pooling) and a combined loss function. The method enhances the local boundary positioning and the contrast of the background foreground by adding the MambaASPP module after the feature extraction stage of the U-shaped model architecture, thereby improving the segmentation accuracy of small blood vessel identification. In addition, a segmentation head is added after each module in the decoding stage to introduce deep supervision and improve the connectivity of vascular segmentation. Finally, the cross entropy, Dice and cbDice (Centerline Boundary Dice) loss functions are combined to optimize the coronary artery segmentation. Finally, the coronary artery is effectively segmented to assist doctors in diagnosing cardiovascular diseases. The present invention refers to the model constructed by the proposed method as AMU-Net.

[0007] A coronary artery segmentation method combining MambaASPP and a combined loss function of the present invention comprises the following steps:

[0008] S1: Data preprocessing stage: perform non-zero data cropping on the image data, extract the minimum bounding box containing the actual foreground, and then resample to adjust the cropped image to the predetermined spatial resolution to ensure the consistency of the data at different scales.

[0009] S2: Feature extraction stage: In the left half of the model U architecture, a residual structure is added to each block in the feature extraction encoder, and then the MambaASPP module is added to the bottleneck layer after feature extraction to enhance the local boundary positioning and the contrast of the background and foreground, thereby improving the segmentation accuracy of small blood vessel identification;

[0010] S3: Feature decoding stage: In the right half of the model U architecture, a segmentation head is added to each layer to obtain the segmentation result. The modules of each layer are supervised through the loss function, so that the model can learn the detailed information in advance;

[0011] S4: Supervised model training, combining cross entropy, Dice and cbDice loss functions with a weight ratio of 3:1:2. The three functions are combined to form a loss function to improve the segmentation effect from different angles.

[0012] Preferably, in step S1, in the cropping stage of data preprocessing, the image data is first cropped in non-zero areas to extract the minimum bounding box containing the actual foreground content. The cropped image data and segmented data only retain the valid foreground area, avoiding the surrounding redundant zero-value parts, thereby reducing the dimension of the data. The shape of the data before cropping and the bounding box of the cropped area will be recorded to ensure that they can be referenced or restored in subsequent processes. In the resampling stage, the image data will be resized according to the target spatial resolution to adapt to the needs of model training. By calculating the target resolution and shape of the cropped data, resampling adjusts the pixel spacing and image size to ensure the consistency of the data at different scales. Resampling is performed after cropping and normalization to prevent data distortion and improve computational efficiency.

[0013] In step S2, the network structure constructed by the present invention is as follows: Figure 1 As shown in the figure, the first layer of the network structure is a 3×3 convolution layer with 32 output channels, which captures the low-level features of the input image and provides rich information for the subsequent network layers. The subsequent layer is a "U"-shaped architecture, with the left half being the feature extraction stage and the right half being the feature decoding stage. Both the feature extraction stage and the decoding stage have five layers.

[0014] In each block of the Encoder, such as Figure 1 As shown in the orange module in the middle, each module contains two convolutional layers, and each convolutional layer has a residual architecture. In addition, there is a normalization layer and activation function after each convolution, and then downsampling is performed through the blue module pooling layer, thereby compressing the amount of data and parameters, reducing overfitting, expanding the receptive field, and continuously extracting high-level semantic information of the image.

[0015] Preferably, in the bottleneck layer after the feature extraction stage, in order to alleviate the problem of resolution degradation after continuous downsampling, which affects the distinction between coronary arteries and other organs, the present invention designs a MambaASPP module. Figure 2As shown in the figure, first, the ASPP module will pass through five parallel branches, including 1×1 convolution kernel, convolution layer with dilation rate of 1, 3×3 convolution kernel, convolution layer with dilation rate of 6, 3×3 convolution kernel, convolution layer with dilation rate of 12, 3×3 convolution kernel, convolution layer with dilation rate of 18, and layer for obtaining global context information, to obtain multi-scale feature information. Then the information is input into Mamba, which acts as a gating mechanism to select the most useful multi-scale features, and then performs dot multiplication with the original multi-scale output, thereby weighting the original output and playing a gating role, enhancing the local boundary positioning and the contrast of the background and foreground, thereby improving the segmentation accuracy of small blood vessel recognition. Finally, the output is obtained through the convolution layer of 1×1 convolution kernel and handed over to the decoding part for processing.

[0016] Preferably, in step S3, the residual structure used in each module of the feature decoding part is consistent with the orange module in the feature extraction stage. After each block of the decoding part outputs a feature map, a segmentation header is added, such as Figure 1 As shown in the purple module in the figure, the segmentation result is obtained. The loss function is used to supervise the modules of each layer, so that the model can learn the detailed information in advance. In general, the architecture of model training is as follows Figure 3 (a), and the training method of the present invention is as follows Figure 3 (b) as shown.

[0017] Preferably, in step S4, based on the above-mentioned model, the present invention uses a loss function combining cross entropy, Dice and cbDice to optimize the effect of coronary artery segmentation and supervise model training.

[0018] The specific formula of the cross entropy loss function is shown in formula (1).

[0019]

[0020] Where p(x) represents the true label and q(x) represents the predicted label. The smaller the cross entropy value, the better the model prediction effect. It measures the accuracy of the model prediction by calculating the distance between the predicted distribution and the true distribution. For pixel-level segmentation tasks, cross entropy can well measure the classification error of each pixel.

[0021] The specific formula of Dice is shown in formula (2).

[0022]

[0023] Among them, V P represents the segmentation mask predicted by the model, V LRepresents the true segmentation mask. The Dice coefficient is calculated by the intersection and union of the predicted segmentation area and the true segmentation area. It is particularly suitable for dealing with problems of imbalanced categories, such as the uneven ratio of blood vessels to background areas in coronary arteries.

[0024] Preferably, the simple use of cross entropy and Dice loss is not sufficient to deal with the connectivity and detail issues in the coronary artery segmentation task. In order to better capture the topological structure and detail information of the coronary artery, the present invention introduces cbDice loss. cbDice loss is an extension of clDice (Centerline Dice). clDice is a novel topology-preserving loss function designed for tubular structure segmentation. It evaluates the connectivity of the segmentation by calculating the intersection of the segmentation mask and their morphological skeletons, and measures the accuracy of the coronary artery centerline segmentation. The specific formula of clDice is shown in Equations (3)-(5).

[0025]

[0026]

[0027]

[0028] Among them, V p represents the segmentation mask predicted by the model, V L represents the true segmentation mask, S P Indicates that from V P The extracted skeleton, S L Indicates that from V L The topologically accurate molecule represents the predicted skeleton S P and the true segmentation label V L The denominator is the predicted skeleton, which represents the accuracy. The numerator represents the actual skeleton S L and the predicted label V P The intersection of , the denominator is the real skeleton. The calculation of clDice combines topological accuracy and topological sensitivity to measure the quality of the segmentation result in a harmonic average manner. It is particularly suitable for tubular structures such as coronary arteries.

[0029] cbDice is an extension of clDice. It optimizes the accuracy and sensitivity calculations and increases the recognition of tubular structure geometry information (such as radius and distance). While maintaining topological integrity, it enhances the focus on small vessels and details by combining boundary perception and radius information extracted from the vascular skeleton. cbDice can be used to evaluate the model's control over the topological structure and overall details of the coronary arteries. The specific formulas of cbDice are shown in equations (6)-(8).

[0030]

[0031]

[0032]

[0033] Among them, as in clDice, V P represents the segmentation mask predicted by the model, V L represents the true segmentation mask, S P Indicates that from V P The extracted skeleton, S L Indicates that from V L The skeleton extracted from P，N Represents the inverse of the predicted skeleton radius, I L,N Represents the inverse of the true skeleton radius, R P，N Represents the predicted skeleton radius, R L,N Indicates the actual skeleton radius, D L,N Indicates the distance from the actual blood vessel point to the blood vessel wall, D P,N Indicates the distance from the predicted blood vessel point to the blood vessel wall. The subscript N indicates that the variable has been normalized. Finally, the topological accuracy and topological sensitivity are combined. By introducing the radius information and the distance from the blood vessel point to the blood vessel wall, the blood vessel structure is focused, and the inverse of the skeleton radius is introduced to make the model focus on the small blood vessels.

[0034] Preferably, the loss function of the supervised model of the present invention combines cross entropy, Dice and cbDice, which can improve the model performance in multiple dimensions. Cross entropy provides basic classification accuracy, Dice enhances the overall segmentation region consistency, and cbDice refines the details and connectivity of the coronary artery structure, so that the model can not only accurately segment the blood vessels, but also maintain the topological structure of the blood vessels. This combination of multiple loss functions can achieve a good balance between accuracy, detail capture and connectivity maintenance, thereby improving the performance of the model in the coronary artery segmentation task.

[0035] In summary, compared with the prior art, the advantages of the present invention are as follows: the present invention uses Mamba as a gating mechanism for the first time, combined with ASPP, and proposes a new architecture AMU-Net. In addition, the present invention designs MambaASPP to fuse the multi-scale information of the bottleneck layer, so as to improve the local boundary positioning and the contrast of the background foreground when processing the abstract contextual semantic information and global information of the bottleneck layer, and assist the Decoder part to better process the global and local information. The present invention also combines the cross entropy, Dice and cbDice loss functions, and achieves a good balance between accuracy, detail capture and connectivity preservation through different loss functions. The present invention also adopts a deep supervision method to allow the model to learn more discriminative features in advance at each layer. Finally, the present invention incorporates clDice and cbDice into the evaluation indicators, focusing on the segmentation effect of vascular connectivity and detail information. Compared with the traditional CNN and Transformer-based methods, and the current Mamba-based methods, the present invention shows a strong ability in maintaining the connectivity of tubular structures such as coronary arteries and segmenting the overall structure. In addition, the present invention is based on the nnUNet architecture and can be directly transplanted to other image segmentation tasks. It has good scalability and is worthy of further in-depth exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The overall network architecture constructed for the present invention;

[0037] Figure 2 The MambaASPP module architecture proposed in the present invention;

[0038] Figure 3 Compare the general model training method and the model training method under deep supervision; DETAILED DESCRIPTION

[0039] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0040] Step 1: Data preprocessing stage: perform non-zero data cropping on the image data, extract the minimum bounding box containing the actual foreground, and then resample to adjust the cropped image to the predetermined spatial resolution to ensure the consistency of the data at different scales;

[0041] Step 2: Feature extraction stage: In the left half of the model U architecture, a residual structure is added to each block in the feature extraction encoder, and then the MambaASPP module is added to the bottleneck layer after feature extraction to enhance the local boundary positioning and the contrast of the background and foreground, thereby improving the segmentation accuracy of small blood vessel recognition;

[0042] Step 3: Feature decoding stage: In the right half of the model U architecture, a segmentation head is added to each layer to obtain the segmentation result. The modules of each layer are supervised through the loss function to allow the model to learn detailed information in advance;

[0043] Step 4: Supervise model training. Combine the cross entropy, Dice, and cbDice loss functions with a weight ratio of 3:1:2. The three functions together constitute the loss function to improve the segmentation effect from different angles.

[0044] The present invention is implemented within the framework of nnUNet because the module design of nnUNet is very consistent with the focus of MambaASPP, the new network architecture we are going to introduce. In addition, nnUNet automatically completes image preprocessing and image enhancement according to the data set, so that the present invention can be compared with other models under unified conditions. Based on nnUNet, the present invention can automatically configure any data set and is suitable for a wide range of image segmentation tasks. Specific implementation method It runs on a single NVIDIA GeForce RTX 3090, trains 250 epochs, has a batch size of 2, and uses five-fold cross validation. Intersection over Union (IoU), Dice coefficient (Dice Similarity Coefficient, DSC), Centerline Dice (clDice) Score, and Centerline Boundary Dice (cbDice) Score are used as evaluation indicators.

[0045] Phase 1: Data preprocessing phase: First, the image data is cropped of non-zero areas to extract the minimum bounding box containing the actual foreground content. The cropped image data and segmentation data only retain the valid foreground area, avoiding the redundant zero-value parts around it, thereby reducing the dimension of the data. The shape of the data before cropping and the bounding box of the cropped area are recorded to ensure that they can be referenced or restored in subsequent processes. In the resampling phase, the image data is resized according to the target spatial resolution to adapt to the needs of model training. By calculating the target resolution and shape of the cropped data, resampling adjusts the pixel spacing and image size to ensure the consistency of the data at different scales. Resampling is performed after cropping and normalization to prevent data distortion and improve computational efficiency.

[0046] Figure 1 The coronary artery segmentation model combining MambaASPP and a combined loss function proposed in the present invention is demonstrated. The structure includes a feature extraction stage, a MambaASPP module for enhancing local boundary positioning and background foreground contrast, improving segmentation accuracy of small vessel identification, and a decoding stage.

[0047] The second stage: feature extraction stage, the data set in the first stage is input for feature extraction. First, a 3×3 convolution layer with an output channel number of 32 is used to capture the low-level features of the input image, and then downsampled through a five-layer feature extraction module to extract relevant image features, retain important image information, and expand the receptive field. Then the MambaASPP module is used to filter the extracted image features. Its structure is as follows: Figure 2 As shown in the figure. First, the extracted image features will pass through five parallel branches, including a 1×1 convolution kernel, a convolution layer with a dilation rate of 1, a 3×3 convolution kernel, a convolution layer with a dilation rate of 6, a 3×3 convolution kernel, a convolution layer with a dilation rate of 12, a 3×3 convolution kernel, a convolution layer with a dilation rate of 18, and a layer for obtaining global context information, to obtain multi-scale feature information. Then, this information is input into Mamba. At this time, Mamba acts as a gating mechanism to filter out the most useful multi-scale features, and then performs a dot multiplication with the original multi-scale output, thereby weighting the original output and playing a gating role, enhancing the contrast between local boundary positioning and background foreground, thereby improving the segmentation accuracy of small blood vessel recognition. Finally, the final output is obtained through a convolution layer with a 1×1 convolution kernel and passed to the decoding part.

[0048] The third stage: feature decoding stage, the features filtered by MambaASPP are input into the five-layer feature decoding modules in turn. A feature map is obtained through each feature decoding module, and then the five feature maps are input into the five segmentation modules respectively, such as Figure 1 As shown in the purple module in the middle, the segmentation prediction results are obtained respectively, so that the modules of each layer are supervised through the loss function, allowing the model to learn detailed information in advance.

[0049] The fourth stage: supervised model training. In the above stage, five segmentation prediction results will be obtained, and then the segmentation label and the segmentation prediction result will be input into the loss function respectively to obtain five model training loss values, and the weighted sum will be obtained to obtain the final loss value. In order to improve the segmentation performance, the loss function of the present invention combines cross entropy, Dice and cbDice according to the weight of 3:1:2 to optimize the effect of coronary artery segmentation and improve the model performance in different dimensions. Cross entropy provides basic classification accuracy, Dice enhances the overall segmentation region consistency, and cbDice refines the details and connectivity of the coronary artery structure, so that the model can not only accurately segment the blood vessels, but also maintain the topological structure of the blood vessels. This combination of multiple loss functions can achieve a good balance between accuracy, detail capture and connectivity maintenance, thereby improving the performance of the model in the coronary artery segmentation task.

[0050] The coronary artery segmentation method provided in the above embodiment in combination with MambaASPP and a combined loss function is first preprocessed by non-zero data clipping and resampling. Then, in the process of training, a series of convolutional layers with a residual structure are used in the feature extraction stage to extract rich contextual semantic information and expand the receptive field. At the same time, considering the problem that the resolution is reduced due to continuous downsampling in the feature extraction stage, which affects the distinction between the coronary artery and other organs, the present invention adopts MambaASPP to enhance the local boundary positioning and the contrast of the background foreground. In the feature decoding stage, a segmentation head is set for each layer, and the loss function is supervised layer by layer to allow the model to learn the detailed information in advance. Finally, a method combining cross entropy, Dice and cbDice is selected to supervise the training to obtain a trained coronary artery segmentation model. The method of the present invention is more effective than other traditional methods, and the detailed information of the coronary artery is comprehensively considered.

[0051] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A coronary artery segmentation method combining MambaASPP and a combined loss function, comprising the following steps: Step 1: Data preprocessing stage: perform non-zero data cropping on the image data, extract the minimum bounding box containing the actual foreground, and then resample to adjust the cropped image to the predetermined spatial resolution to ensure the consistency of the data at different scales; Step 2: Feature extraction stage: In the left half of the model U architecture, a residual structure is added to each block in the feature extraction encoder, and then the MambaASPP module is added to the bottleneck layer after feature extraction to enhance the local boundary positioning and the contrast of the background and foreground, thereby improving the segmentation accuracy of small blood vessel recognition; Step 3: Feature decoding stage: In the right half of the model U architecture, a segmentation head is added to each layer to obtain the segmentation result. The modules of each layer are supervised through the loss function to allow the model to learn detailed information in advance. Step 4: Supervise model training. Combine the cross entropy, Dice, and cbDice loss functions with a weight ratio of 3:1:

2. The three functions together constitute the loss function to improve the segmentation effect from different angles.

2. The method for coronary artery segmentation combining MambaASPP and a combined loss function according to claim 1, characterized in that: In the cropping stage of data preprocessing, the non-zero area of ​​the image data is first cropped to extract the minimum bounding box containing the actual foreground content. The cropped image data and segmentation data only retain the valid foreground area, avoiding the redundant zero-value parts around it, thereby reducing the dimension of the data. The shape of the data before cropping and the bounding box of the cropped area will be recorded to ensure that they can be referenced or restored in subsequent processes. In the resampling stage, the image data will be resized according to the target spatial resolution to adapt to the needs of model training. By calculating the target resolution and shape of the cropped data, resampling adjusts the pixel spacing and image size to ensure the consistency of the data at different scales. Resampling is performed after cropping and normalization to prevent data distortion and improve computational efficiency.

3. The method for coronary artery segmentation combining MambaASPP and a combined loss function according to claim 1, characterized in that: The first layer of the network structure is a 3×3 convolution layer with 32 output channels, which captures the low-level features of the input image and provides rich information for the subsequent network layers. The subsequent layer is a "U"-shaped architecture, with the left half being the feature extraction stage and the right half being the feature decoding stage. Both the feature extraction stage and the decoding stage have five layers. In each block of the feature extraction stage, as shown in the orange module in Figure 1, each module contains two convolution layers, and each convolution layer has a residual architecture. In addition, there is a normalization layer and an activation function after each convolution, and then downsampling is performed through the blue module pooling layer, thereby compressing the amount of data and parameters, reducing overfitting, retaining important image features, and expanding the receptive field. In the bottleneck layer after the feature extraction stage, the MambaASPP module is designed. First, AS The PP module will obtain multi-scale feature information through five parallel branches, namely, a 1×1 convolution kernel, a convolution layer with a dilation rate of 1, a 3×3 convolution kernel, a convolution layer with a dilation rate of 6, a 3×3 convolution kernel, a convolution layer with a dilation rate of 12, a 3×3 convolution kernel, a convolution layer with a dilation rate of 18, and a layer for obtaining global context information. This information is then input into Mamba, where Mamba acts as a gating mechanism to screen out the most useful multi-scale features, and then performs a dot multiplication with the original multi-scale feature output to weight the original output, thereby playing a gating role and enhancing the local boundary positioning and the contrast of the background foreground, thereby improving the segmentation accuracy of small blood vessel recognition. Finally, the output is obtained through a convolution layer with a 1×1 convolution kernel and handed over to the decoding part for processing.

4. The method for coronary artery segmentation combining MambaASPP and a combined loss function according to claim 1, characterized in that: The residual structure used in each module of the feature decoding part is consistent with the orange module in the feature extraction stage. After each output of the decoding part obtains the feature map, a segmentation head is added, as shown in the purple module in Figure 1, to obtain the segmentation result. The loss function is used to supervise the modules of each layer, so that the model can learn the detailed information in advance.

5. The coronary artery segmentation method according to claim 1, wherein During model training, a loss function combining cross entropy, Dice and cbDice was used to optimize the effect of coronary artery segmentation and supervise model training. The specific formula of the cross entropy loss function is shown in formula (1): Where p(x) represents the true label, q(x) represents the predicted label, and the specific formula of Dice is shown in formula (2): Among them, V p represents the segmentation mask predicted by the model, V L represents the true segmentation mask. The cbDice loss is an extension of the clDice (Centerline Dice). The specific formula of clDice is shown in equations (3)-(5): Among them, V P represents the segmentation mask predicted by the model, V L represents the true segmentation mask, S P Indicates that from V P The extracted skeleton, S L Indicates that from V L The extracted skeleton, Tprec represents the topological accuracy, and its numerator represents the predicted skeleton S P and the true segmentation label V L The denominator is the predicted skeleton, which represents the accuracy. Tsens represents the topological sensitivity, and its numerator represents the actual skeleton S L and the predicted label V P The denominator is the real skeleton. The calculation of clDice combines topological accuracy and topological sensitivity to measure the quality of the segmentation result in a harmonic average manner. It is particularly suitable for tubular structures such as coronary arteries. The specific formula of cbDice is shown in equations (6)-(8): Among them, as in clDice, V P represents the segmentation mask predicted by the model, V L represents the true segmentation mask, S P Indicates that from V P The extracted skeleton, S L Indicates that from V L The skeleton extracted from P,N Represents the inverse of the predicted skeleton radius, I L,N Represents the inverse of the true skeleton radius, R P,N Represents the predicted skeleton radius, R L,N Indicates the actual skeleton radius, D L,N Indicates the distance from the actual blood vessel point to the blood vessel wall, D P,N Represents the distance from the predicted blood vessel point to the blood vessel wall, where the subscript N indicates that the variable has been normalized. Finally, the topological accuracy and topological sensitivity are also combined. By introducing the radius information and the distance from the blood vessel point to the blood vessel wall, the blood vessel structure is focused on, and the inverse of the skeleton radius is introduced to make the model focus on the small blood vessel part. In the model, the loss function of the supervised model combines the above-mentioned cross entropy, Dice and cbDice, which can improve the model performance in multiple dimensions.

Images