Double-view tubular tissue segmentation method based on template guidance
Through template conditions, the deformable convolution and cross-fusion attention module, combined with the residual self-attention feature fusion strategy, the existing methods are solved inadequate morphological information extraction and feature fusion in tubular tissue segmentation, achieving higher accuracy and robust segmentation effects.
Patent Information
- Application Number
- CN202510478974.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-16
- Publication Date
- 2025-08-12
AI Technical Summary
In the segmentation of tubular tissue, existing methods have problems such as over-curing of templates, insufficient fusion of global and local features, lack of general models, insufficient extraction of morphological information, and excessively solidified morphological information extraction methods, resulting in inaccurate and discontinuous segmentation results.
The template-guided dual-view tubular tissue segmentation method is adopted to extract geometric and topological information of tubular tissue through template conditional deformable convolution (TCDConv), cross-fusion attention and convolution (CCFA) modules, and residual self-attention feature fusion strategy (RSA), and image resolution is restored through layer by layer upsampling and feature fusion.
It significantly improves the accuracy and robustness of tubular tissue segmentation, especially the segmentation effect in complex backgrounds, and enhances the model's adaptability and generalization ability to different tubular tissues.
Smart Images
Figure CN120472154A_ABST
Abstract
Description
Technical Field
[0001] The present invention mainly relates to the technical field of medical image processing, and in particular to a dual-view tubular tissue segmentation method based on template guidance. Background Art
[0002] Medical image segmentation is a key step in computer-aided diagnosis (CAD), which aims to extract anatomical tissues of interest from medical images. The segmentation of tubular tissues is particularly challenging because these tissues often have complex morphologies and topological relationships. In the field of medical image processing technology, accurate segmentation of tubular tissues (such as blood vessels, colorectum, bronchi, etc.) is crucial for disease diagnosis, treatment planning, and prognosis assessment. Traditional tubular tissue segmentation methods mainly rely on manually designed features such as texture, intensity, and shape. However, existing traditional methods perform poorly when dealing with complex backgrounds and have difficulty coping with noise, low contrast, and complex anatomical tissues in medical images.
[0003] In recent years, deep learning-based methods for tubular tissue segmentation have made significant progress. U-Net and its variants (such as 3D U-Net and nnUNet), through their symmetrical encoder-decoder architecture, effectively capture image context and perform accurate segmentation. Furthermore, improved network architectures (such as Swin-UNet and TransUNet) have further enhanced global feature extraction capabilities by incorporating the Transformer mechanism. However, existing traditional methods often neglect the extraction of morphological information when processing tubular tissue, resulting in inaccurate segmentation results, particularly when dealing with curved and narrow tubular tissue.
[0004] Deformable convolution can adaptively adjust the sampling position of the convolution kernel by introducing a learnable offset, thereby better capturing the geometric and topological features of the image. For example, Dynamic Snake Convolution (DSConv) can more accurately capture the morphological characteristics of specific tissues by adaptively focusing on local information of curvature and narrowness.
[0005] In existing traditional technologies, the general process of tubular tissue segmentation based on deformable convolution includes: ① Input image preprocessing: The input 3D medical image (such as CT or MRI) is preprocessed, including normalization, cropping, and resampling operations to ensure the uniformity of the input data.
[0006] ② Feature extraction: Deformable convolution layers are used to extract geometric and topological features of the image. By introducing a learnable offset, deformable convolution can adaptively adjust the sampling position of the convolution kernel to better capture the morphological information of tubular tissue.
[0007] ③ Feature fusion: The extracted local features are fused with the global features, usually through skip connections or attention mechanisms. This step aims to balance local details and overall organizational information.
[0008] ④ Upsampling and segmentation: Upsampling gradually restores the image resolution and generates the final segmentation result. Common upsampling methods include deconvolution or interpolation.
[0009] ⑤ Loss calculation and optimization: Use loss functions such as Dice coefficient and cross entropy to calculate the difference between the segmentation result and the true label, and optimize the network parameters through the back propagation algorithm.
[0010] However, the existing traditional deformable convolution method has the following shortcomings: (1) The template is too rigid: Existing deformable convolution templates are usually too rigid and difficult to flexibly adapt to the morphological characteristics of different tubular tissues, resulting in insufficient morphological information extraction.
[0011] (2) Insufficient fusion of global and local features: Existing methods perform poorly in fusing global and local features, resulting in inaccurate segmentation results in both details and overall organization.
[0012] (3) Lack of a universal model: Existing methods usually design network architectures for specific tubular tissues and lack a universal model that can handle a variety of tubular tissues.
[0013] (4) Insufficient extraction of morphological information: Existing methods mainly focus on pixel-level information and ignore the extraction of morphological information, resulting in discontinuous segmentation results.
[0014] (5) Morphological information extraction methods are too rigid: Existing morphological information extraction techniques (such as deformable convolution) are too rigid and cannot flexibly capture the geometric and topological features of tubular tissues. Summary of the Invention
[0015] The technical problem to be solved by the present invention is that, in response to the above-mentioned problems of the prior art, the present invention provides a template-guided dual-view tubular tissue segmentation method with simple principle, convenient operation, high stability and good applicability.
[0016] In order to solve the above technical problems, the technical solution adopted by the present invention is: A dual-view tubular tissue segmentation method based on template guidance, comprising: Step S1: Extracting the geometric and topological information of tubular tissue through template-conditional deformable convolution; Step S2: extract global, key and local features; Step S3: gradually restore the image resolution through layer-by-layer upsampling and feature fusion.
[0017] As a further improvement of the present invention: Step S1 includes: Step S11: Determine the optimal offset through morphological offset clustering algorithm ; Step S12: adjusting the receptive field of the convolution kernel according to the offset to extract morphological information; Step S13: Design conditional convolution according to the results of the morphological shift clustering algorithm.
[0018] As a further improvement of the present invention: in step S11, by measuring The degree of deviation is used to calculate the routing weight of the example dependency , the process of determining the routing weight is as follows: ; in, Represents the position of the center point of the image, is the threshold hyperparameter.
[0019] As a further improvement of the present invention: in step S2, global, key and local features are extracted through a cross-fusion module of attention mechanism and convolution; the cross-fusion module of attention mechanism and convolution includes three paths: a convolution path, a self-attention path and an attention path.
[0020] As a further improvement of the present invention: the convolution path includes shifting and aggregating operations. The volume integral is decomposed into indivual Convolution, extracting local detail information:
[0021] in, represents the center position of the convolution kernel, Represents the size of the convolution kernel, It's ordinary Convolution, and Indicates the global offset.
[0022] As a further improvement of the present invention: the self-attention path includes: implementing a window shift mechanism by expanding and stacking operations, The dimensions are decomposed into their square roots along the width and height dimensions of the image , the process of performing self-attention calculation during window shifting is expressed as: .
[0023] As a further improvement of the present invention: the attention path includes extracting key information through a convolutional block attention module (CBAM):
[0024] in, represents the spatial attention module, represents the channel attention module, Stands for element-wise multiplication.
[0025] As a further improvement of the present invention: the outputs of the attention mechanism and the convolution cross-fusion module are fused through cross-self-attention calculation, and the outputs of the two paths are input into the cross-self-attention calculation process.
[0026] As a further improvement of the present invention: the attention branch generates the original 、 and , these are generated by the self-attention path; and is obtained in the same way through the convolution path; two cross-attention calculation operations are performed on the similarity, where Represents similarity calculation, is a depthwise convolution; its formula is as follows: .
[0027] As a further improvement of the present invention: adopting the residual self-attention feature fusion strategy, the formula is as follows:
[0028]
[0029] in, It represents the average value after adding the elements. Represents an upsampling operation.
[0030] Compared with the prior art, the advantages of the present invention are: The present invention's template-guided, dual-view tubular tissue segmentation method significantly improves the accuracy and robustness of tubular tissue segmentation by introducing template conditional deformable convolution (TCDConv), a cross-fused attention and convolution (CCFA) module, and a residual self-attention feature fusion strategy (RSA). First, TCDConv dynamically adjusts the receptive field of the convolution kernel to more accurately capture the geometric and topological features of tubular tissue, avoiding the information loss problem of traditional convolution kernels in complex backgrounds. Second, the CCFA module effectively integrates global, local, and key features through multi-path feature extraction and cross-self-attention computation, enhancing the model's adaptability to complex tubular tissues. Finally, the RSA strategy further enhances feature fusion through residual and self-attention mechanisms, avoids the vanishing gradient problem, and ensures the model's generalization across diverse datasets. Experimental results demonstrate that the proposed method achieves superior segmentation results compared to existing methods on multiple public datasets, particularly with significantly improved segmentation accuracy and robustness in complex backgrounds.
[0031] In summary, the present invention solves the shortcomings of existing methods in morphological information extraction and feature fusion through innovative convolution and attention mechanisms, significantly improves the accuracy and robustness of tubular tissue segmentation, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 It is a schematic diagram of the process of the present invention in a specific application example.
[0032] Figure 2 It is a schematic diagram of information extraction through template-conditional deformable convolution in a specific application example of the present invention.
[0033] Figure 3 It is a schematic diagram of the principle of the overall network architecture in a specific embodiment of the present invention.
[0034] Figure 4 It is a schematic diagram of an implementation method of template-conditional deformable convolution in a specific embodiment of the present invention. DETAILED DESCRIPTION
[0035] The present invention will be further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the scope of protection of the present invention is not limited thereby.
[0036] In the description of this invention, it is necessary to understand some professional terms. The following are explanations of the key terms: (1) Tubular Structure: Tubular tissue refers to anatomical tissue with a tubular morphology in the human body, such as blood vessels, intestines, and bronchi. These tissues typically appear as elongated, complexly branched structures in medical images, and their accurate segmentation is crucial for disease diagnosis and treatment.
[0037] (2) Morphological Features: Morphological features refer to geometric information such as the shape and topological organization of objects in an image. In medical image processing, morphological features are often used to describe the curvature, branching, thickness, and other characteristics of tubular tissue.
[0038] (3) Pixel-level Features: Pixel-level features refer to the intensity, color, and other information of each pixel in an image. In medical image segmentation, pixel-level features are often used to distinguish the boundaries of different tissues or organs.
[0039] (4) Template Conditional Deformable Convolution (TCDConv): TCDConv is an improved convolution operation that introduces learnable offsets to adjust the sampling position of the convolution kernel, thereby better capturing the morphological features in the image. Unlike traditional fixed convolution kernels, TCDConv can dynamically adjust the shape of the convolution kernel based on the image content, making it suitable for processing complex tubular tissues.
[0040] (5) Self-Attention Mechanism The self-attention mechanism is a technique for capturing long-range dependencies in images. It extracts global features by calculating correlations between different locations in the image. In medical image segmentation, the self-attention mechanism is often used to handle complex backgrounds and tissues. The multi-head self-attention mechanism is an extension of the self-attention mechanism, capturing different features in the image by processing multiple attention heads in parallel. The multi-head self-attention mechanism can more comprehensively extract global information and is suitable for processing complex medical images.
[0041] (5) Crossing Fusion of Attention and Convolution (CCFA): CCFA is a feature fusion module that combines the advantages of the self-attention mechanism and convolution operations. By extracting global features, key features, and local features through three independent paths, CCFA can more comprehensively capture the information in the image, thereby improving segmentation accuracy.
[0042] (6) Residual Self-Attention Feature FusionStrategy (RSA): RSA is a feature fusion strategy that fuses features from different layers through a self-attention mechanism and residual connections. This strategy can effectively avoid the vanishing gradient problem and enhance the model's feature fusion capabilities.
[0043] (7) Trilinear Interpolation: Trilinear interpolation is a method for upsampling three-dimensional images, estimating the pixel value of an unknown point by linearly interpolating the three dimensions. In medical image processing, trilinear interpolation is often used to restore low-resolution images to high resolution.
[0044] (8) Dice Coefficient: The Dice coefficient is a metric used to evaluate image segmentation accuracy. It calculates the degree of overlap between the predicted segmentation result and the true label. The Dice coefficient ranges from 0 to 1, with higher values indicating more accurate segmentation results.
[0045] (9) Hausdorff Distance: The Hausdorff distance is an indicator used to evaluate the edge accuracy of the segmentation result. It calculates the maximum distance between the predicted segmentation result and the true label. The smaller the Hausdorff distance, the more accurate the edge of the segmentation result.
[0046] (10) Centerline Dice coefficient (clDice): clDice is a metric specifically designed to assess the accuracy of tubular tissue segmentation. It evaluates segmentation quality by calculating the degree of overlap between the predicted segmentation result and the centerline of the true label. clDice can better reflect the topological integrity of tubular tissue.
[0047] Example like Figures 1-4 As shown, the template-guided dual-view tubular tissue segmentation method of the present invention includes the following steps: Step S1: Extract the geometric and topological information of tubular tissue through template conditional deformable convolution (TCDConv); Step S2: extract global, key and local features; Step S3: gradually restore the image resolution through layer-by-layer upsampling and feature fusion.
[0048] In a specific embodiment, the formula of the above deformable convolution (TCDConv) is as follows:
[0049] in, represents the position in the input and output features, is the first in the convolution sampling grid positions, is the corresponding offset, It is a sample-dependent scalar weight used to determine whether to apply an offset. Conventional convolution is considered to be weak in focusing on the morphological information of distant areas, and it is not desirable to over-shift the receptive field, which may lead to the loss of morphological features in the proximal area. Therefore, this paper introduces conditional deformation offset. .in Indicates that it corresponds to The convolution sampling grid The center position is offset, and Indicates that a routing function is used to calculate the The scalar weight of the associated instance dependency, used to determine whether to apply a bias.
[0050] See also Figure 2 In a specific application example, step S1 further includes: Step S11: Determine the optimal offset through morphological offset clustering algorithm ; Step S12: adjusting the receptive field of the convolution kernel according to the offset to extract morphological information; Step S13: Design conditional convolution according to the results of the morphological shift clustering algorithm.
[0051] Furthermore, in step S11, by measuring The degree of deviation is used to calculate the routing weight of the example dependency , the process of determining the routing weight is as follows: ; in, Represents the position of the center point of the image, is the threshold hyperparameter.
[0052] In a specific application example, the principle of the morphological shift clustering algorithm adopted by the present invention is as follows: 1. For each image in the training set, get the position of each pixel , Strength value , curvature .
[0053] 2. Use The algorithm clusters the multidimensional vectors and obtains the cluster center of each image.
[0054] 3. Cluster the centers of all images again Clustering to get the final offset .
[0055] In a specific application example, in step S2, global, key and local features are extracted through the attention mechanism and convolution cross fusion module (CCFA).
[0056] The cross-fusion module of attention mechanism and convolution (CCFA) includes three paths: convolution path, self-attention path and attention path.
[0057] Furthermore, the convolution path includes offset operations and aggregation operations to convert The volume integral is decomposed into indivual Convolution, extracting local detail information:
[0058] in, represents the center position of the convolution kernel, Represents the size of the convolution kernel, It's ordinary Convolution, and Indicates the global offset.
[0059] Furthermore, the self-attention path includes a window shifting mechanism implemented through expansion and stacking operations to reduce computational complexity and focus on important regions. This window shifting mechanism is implemented using expansion and stacking operations to reduce computational complexity and focus on important regions within a larger context. The expansion operation evenly spreads the depth dimension of a 3D human tubular tissue image across the height and width of a 2D plane, while the stacking operation is the inverse of the expansion operation.
[0060] Specifically, The dimensions are decomposed into their square roots along the width and height dimensions of the image , the process of performing self-attention calculation during window shifting can be expressed as: .
[0061] Furthermore, the attention path includes extracting key information through the Convolutional Block Attention Module (CBAM):
[0062] in, represents the spatial attention module, represents the channel attention module, Stands for element-wise multiplication.
[0063] As can be seen from the above, the outputs of the attention mechanism and the convolution cross fusion module (CCFA) in the present invention are fused through cross self-attention calculation, and the outputs of the two paths are input into the cross self-attention calculation process. The attention branch generates the original 、 and , these are generated by the self-attention path; and Then it is obtained in the same way through the convolution path. Two cross-attention calculation operations are performed on the similarity. Represents similarity calculation, is the depthwise convolution; the formula is as follows: .
[0064] In this embodiment, the formula of the residual self-attention feature fusion strategy is as follows:
[0065]
[0066] in, It represents the average value after adding the elements. Represents the upsampling operation. The residual self-attention feature fusion strategy can effectively fuse morphological information and pixel-level information, avoid gradient vanishing, and enhance the model's unification ability.
[0067] The core of this invention is a template-guided dual-view tubular tissue segmentation network (TGDNet), which aims to achieve accurate segmentation of various human tubular tissues by fusing morphological information and pixel-level information. Figure 3 For example, the overall architecture principle diagram of TGDNet is shown below, which mainly includes the following modules: (1) Morphological downsampling branch: The geometric and topological information of tubular tissue is extracted through template conditional deformable convolution (TCDConv).
[0068] (2) Pixel feature downsampling branch: used to extract global, key and local features.
[0069] (3) Residual self-attention feature fusion strategy (RSA): gradually restore the image resolution through layer-by-layer upsampling and feature fusion.
[0070] The input , are directed to the morphological downsampling and pixel-level downsampling branches in the network. Morphological downsampling uses template-conditional deformable convolution to extract geometric and topological information , while the pixel-level downsampling branch uses the cross-fusion of attention and convolution to refine the global features, key features and local features from the image Then, the morphological and pixel features extracted by the two downsampling branches are The feature fusion strategy is combined. In the bottleneck block of our model, and are concatenated to obtain the original input of the upsampled part Finally, by trilinear interpolation and Feature fusion strategy layer-by-layer decoding ,in represents the number of layers in our template-guided two-view fusion network.
[0071] 4.2 Morphological Downsampling Branch The main function of the morphological downsampling branch is to extract the morphological information of tubular tissues through template conditional deformable convolution (TCDConv). The core idea of TCDConv is to dynamically adjust the receptive field of the convolution kernel by introducing a conditional deformable offset, thereby more accurately capturing the geometric and topological characteristics of tubular tissues.
[0072] 4.2.1 Template Conditioned Deformable Convolution The formula for TCDConv is as follows:
[0073] in, represents the position in the input and output features, is the first in the convolution sampling grid positions, is the corresponding offset, Is a sample-dependent scalar weight used to decide whether to apply an offset. Conventional convolution is considered to be weak in focusing on the morphological information of distant areas. We do not want the receptive field to be excessively offset, resulting in the loss of morphological features in the proximal area. Therefore, we introduce conditional deformation offset .in Indicates that it corresponds to The convolution sampling grid The center position is offset, and Indicates that a routing function is used to calculate the The scalar weight of the associated instance dependency, used to determine whether to apply a bias.
[0074] The implementation process of TCDConv is as follows Figure 2First, the optimal offset is determined by the Morphological Offset Clustering Algorithm. , and then adjust the receptive field of the convolution kernel according to the offset to extract more accurate morphological information. Then we designed a conditional convolution based on the results of the morphological offset clustering algorithm. The following equation describes its operation mechanism, where represents the magnitude of a vector, also known as the Euclidean distance between two points, Represents the position of the center point of the image, is a threshold hyperparameter. We measure The degree of deviation is used to calculate the routing weight of the example dependency ,The process of determining the routing weight is shown in the following formula.
[0075]
[0076] The template-guided dual-view fusion network (TGDNet) proposed in this paper addresses the shortcomings of existing methods in morphological information extraction and feature fusion by introducing template conditional deformable convolution (TCDConv) and cross-fusion attention and convolution module (CCFA). The specific differences are as follows: Template-conditional deformable convolution: By introducing template-conditional deformable convolution, the present invention can flexibly adjust the sampling position of the convolution kernel, thereby better capturing the morphological information of tubular tissue.
[0077] Dual-view feature fusion: This paper uses a dual-view feature fusion strategy to effectively fuse morphological features and pixel-level features, thereby improving the accuracy and continuity of segmentation results.
[0078] General model: This paper designs a general template-conditional deformable convolution, which can effectively extract the morphological features of different types of tubular tissues and has strong generalization ability.
[0079] It can be seen from the above that in the above scheme of the present invention, the accuracy and robustness of tubular tissue segmentation are significantly improved by introducing template conditional deformable convolution (TCDConv), cross fusion module of attention mechanism and convolution (CCFA), and residual self-attention feature fusion strategy (RSA).
[0080] First, TCDConv dynamically adjusts the receptive field of the convolution kernel to more accurately capture the geometric and topological features of tubular tissues, avoiding the information loss problem of traditional convolution kernels in complex backgrounds. Second, the CCFA module effectively integrates global, local, and key features through multi-path feature extraction and cross-self-attention calculation, enhancing the model's adaptability to complex tubular tissues. Finally, the RSA strategy further enhances the effect of feature fusion through residual and self-attention mechanisms, avoiding the vanishing gradient problem and ensuring the model's generalization ability across different datasets.
[0081] Experimental results show that the present invention has achieved better segmentation results than existing methods on multiple public datasets, especially in complex backgrounds, with significantly improved segmentation accuracy and robustness. For example, on the RAOS dataset, the mDice (average Dice coefficient) of the present invention reached 94.08%, an improvement of 5.43% over the baseline model, while the average Hausdorff distance reached 1.486, an improvement of 4.13% over the baseline model. On all datasets used in the experiment, the average improvement of these two indicators was 6.93% and 10.61%, respectively. In addition, the present application also verified the effectiveness of the TCDConv, CCFA, and RSA modules through ablation experiments, among which TCDConv had the strongest improvement effect and the RSA module had a relatively weaker improvement effect, further demonstrating the superiority of the present invention in the task of tubular tissue segmentation.
[0082] In summary, the present invention solves the shortcomings of existing methods in morphological information extraction and feature fusion through innovative convolution and attention mechanisms, significantly improves the accuracy and robustness of tubular tissue segmentation, and has broad application prospects.
[0083] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A dual-view tubular tissue segmentation method based on template guidance, characterized in that: include: Step S1: Extracting the geometric and topological information of tubular tissue through template-conditional deformable convolution; Step S2: extract global, key and local features; Step S3: gradually restore the image resolution through layer-by-layer upsampling and feature fusion.
2. The template-guided dual-view tubular tissue segmentation method according to claim 1, characterized in that: The step S1 comprises: Step S11: Determine the optimal offset through morphological offset clustering algorithm ; Step S12: adjusting the receptive field of the convolution kernel according to the offset to extract morphological information; Step S13: Design conditional convolution according to the results of the morphological shift clustering algorithm.
3. The template-guided dual-view tubular tissue segmentation method according to claim 2, characterized in that: In step S11, by measuring The degree of deviation is used to calculate the routing weight of the example dependency , the process of determining the routing weight is as follows: ； in, Represents the position of the center point of the image, is the threshold hyperparameter.
4. The template-guided dual-view tubular tissue segmentation method according to any one of claims 1 to 3, characterized in that: In step S2, global, key and local features are extracted through a cross-fusion module of attention mechanism and convolution; the cross-fusion module of attention mechanism and convolution includes three paths: convolution path, self-attention path and attention path.
5. The template-guided dual-view tubular tissue segmentation method according to claim 4, characterized in that: The convolution path includes offset operation and aggregation operation. The volume integral is decomposed into indivual Convolution, extracting local detail information: in, represents the center position of the convolution kernel, Represents the size of the convolution kernel, It's ordinary Convolution, and Indicates the global offset.
6. The template-guided dual-view tubular tissue segmentation method according to claim 4, characterized in that: The self-attention path includes: implementing a window shift mechanism through expansion operations and stacking operations, The dimensions are decomposed into their square roots along the width and height dimensions of the image , the process of performing self-attention calculation during window shifting is expressed as: 。 7. The template-guided dual-view tubular tissue segmentation method according to claim 4, characterized in that: The attention path includes extracting key information through the convolutional block attention module CBAM: in, represents the spatial attention module, represents the channel attention module, Stands for element-wise multiplication.
8. The template-guided dual-view tubular tissue segmentation method according to claim 4, characterized in that: The outputs of the attention mechanism and the convolution cross-fusion module are fused through cross-self-attention calculation, and the outputs of the two paths are input into the cross-self-attention calculation process.
9. The template-guided dual-view tubular tissue segmentation method according to claim 8, characterized in that: The attention branch generates the original 、 and , these are generated by the self-attention path; and It is obtained in the same way through the convolution path; Two cross-attention calculation operations are performed on the similarity, where Represents similarity calculation, It is a depthwise convolution; The formula is as follows: 。 10. The template-guided dual-view tubular tissue segmentation method according to claim 9, characterized in that: The residual self-attention feature fusion strategy is adopted, and its formula is as follows: in, It represents the average value after adding the elements. Represents an upsampling operation.