A blood vessel image segmentation method based on visual manhattan context perception semantics

Abstract

This invention discloses a blood vessel image segmentation method based on Visual Mamba context-aware semantics. It uses a U-shaped encoder-decoder network model for blood vessel image segmentation. The encoder is built around a CSVM module, which is an improved VSS module structure. Specifically, the DW convolutional layer in VMamba's VSS module is replaced by a CCSA module, thus displaying and preserving local and global dependencies in a compressed form. A multi-scale edge guidance module is set between the encoder and decoder, using the Laplacian operator to emphasize the boundary features of low-level features. Simultaneously, a dynamic boundary awareness module is set between the encoder, the multi-scale edge guidance module, and the decoder to aggregate the boundary features of low-level features and the semantic information of high-level features.

Description

Technical Field

[0001] This invention relates to the field of vascular image processing technology, specifically to a vascular image segmentation method based on visual Mamba context-aware semantics. Background Technology

[0002] Early diagnosis and medical intervention are crucial for improving patients' health and prognosis in vascular diseases. Clinicians can use vascular imaging techniques to observe abnormalities in blood vessels and intervene in the early stages of disease. However, sufficient clinical experience is required for accurate diagnosis. Deep learning (DL) technology, which effectively utilizes data, has opened up important new research avenues for medical image analysis.

[0003] Traditionally, deep learning models used for image segmentation tasks have generally been based on U-shaped encoder-decoder networks, with the encoder typically employing a CNN or Vision Transformer. "U-Net: Convolutional Networks for Biomedical Image Segmentation" points out that UNet is a convolutional neural network (CNN) architecture for medical image segmentation. Its excellent segmentation performance and structural simplicity have made it a classic model in image segmentation tasks. UNet is primarily used for pixel-level classification tasks, especially suitable for scenarios requiring precise boundary segmentation, such as medical image analysis. UNet consists of two main parts: an encoder (downsampling path) and a decoder (upsampling path), as well as skip connections between them.

[0004] In a convolutional neural network, each convolutional layer operates on blood vessel images through small convolutional kernels (such as 3x3 or 5x5). Although small convolutional kernels can effectively capture local features, they also limit the receptive field. That is, each convolutional kernel can only focus on a local area of ​​the image and cannot directly capture a larger range of global information.

[0005] While the Vision Transformer can capture global image information, it is limited by the attention mechanism, especially when modeling long sequences, resulting in high quadratic complexity. This leads to expensive computational overhead when handling downstream intensive prediction tasks (such as object detection and semantic segmentation). Furthermore, in medical imaging, high-resolution images such as whole-scale pathological images are common, where transformer-based models may suffer from overfitting and require large amounts of data for effective training.

[0006] Current networks based on Visual Mamba (VMamba) can achieve a global receptive field similar to that of a visual transformer while maintaining linear complexity in terms of token count. However, due to its sequential nature, existing VMamba models still struggle to maintain spatial locality and global dependencies of tokens in high-dimensional arrays.

[0007] In summary, although current encoder-decoder-based U-shaped network structures perform well in most medical image segmentation tasks, they have certain limitations in vascular image segmentation. The main reasons are: vascular structures exhibit significant morphological variations, including thick vessels, thin vessels, and filamentous retinal vessels; small vessels have low contrast, diverse shapes, and are generally less than 10 pixels in length, some even less than 1 pixel; non-vascular structures or lesions (such as the optic disc) may affect segmentation performance; and artifacts generated due to improper data acquisition cannot accurately reconstruct vascular images. Summary of the Invention

[0008] The present invention aims to solve the aforementioned technical problems existing in the prior art by providing a blood vessel image segmentation method based on visual Mamba context-aware semantics.

[0009] The technical solution of this invention is: a blood vessel image segmentation method based on visual Mamba context-aware semantics. This method uses a U-shaped encoder-decoder network model for blood vessel image segmentation. The encoder is built around a CSVM module, which is an improved VSS Block structure. Specifically, the DW convolutional layer before SS2D in the VSS Block is replaced by a CCSA module, thus displaying and preserving local and global dependencies in a compressed form. A multi-scale edge guidance module is set between the encoder and decoder, using the Laplacian operator to emphasize the boundary features of low-level features. Simultaneously, a dynamic boundary awareness module is set between the encoder, the multi-scale edge guidance module, and the decoder to aggregate the boundary features of low-level features and the semantic information of high-level features.

[0010] The encoder, from top to bottom, consists of a residual block ResBlock1-1, a downsampling module Downsampling1-1, a CSVM module 1-1, a downsampling module Downsampling1-2, a CSVM module 1-2, a downsampling module Downsampling1-3, and a CSVM module 1-3. The CCSA consists of a channel attention module ChannelAttention, a channel prior module ChannelPrior, two average pooling layers XAvgPool and YAvgPool, a one-dimensional convolutional layer MS-DW Conv with kernel sizes of 3, 5, 7, and 9, two concat function layers, a group normalization layer Group Norn, and a ReLU activation function layer.

[0011] The decoder consists of, from bottom to top, a residual block ResBlock2-1, an upsampling module Up sampling2-1, a residual block ResBlock2-2, an upsampling module Up sampling2-2, a residual block ResBlock2-3, an upsampling module Upsampling2-3, a residual block ResBlock2-4, and a Seg Head;

[0012] The residual block ResBlock1-1 and the residual block ResBlock2-4 are connected through the multi-scale edge guidance module EGAA1-1; the CSVM module 1-1 and the upsampling module Upsampling2-3 are connected through the multi-scale edge guidance module EGAA1-2; and the CSVM module 1-2 and the upsampling module Upsampling2-2 are connected through the multi-scale edge guidance module EGAA1-3.

[0013] The CSVM module 1-3 and the residual block ResBlock 2-1 are connected by one jump connection and the other is connected through the dynamic boundary awareness module DBA.

[0014] The multi-scale edge guidance modules EGAA1-1, EGAA1-2, and EGAA1-3 have the same structure, and are provided with parallel reverse operation modules Reverse1, Gaussian filters GF, and deep convolutional layers DWConv. The outputs of the reverse operation modules Reverse1, Gaussian filters GF, and deep convolutional layers DWConv pass through convolutional layer Conv1 and are then connected to parallel convolutional layers Conv2, Conv3, and Conv4.

[0015] The Dynamic Boundary Awareness Module (DBA) is equipped with a Dynamic Filter. The output of the Dynamic Filter is divided into two paths: one path consists of a linear layer (Linear layer 1-1), a normalization layer (LayerNorm 1-1), a convolutional layer (Conv 5), an activation function (Sigmoid 1), and a reverse operation module (Reverse 2) in sequence; the other path consists of a linear layer (Linear layer 2-1), a normalization layer (LayerNorm 2-1), a convolutional layer (Conv 6), and an activation function (Sigmoid 2) in sequence. The output features are then passed through the two paths and connected to the Softmax activation function layer. One path consists of a linear layer (Linear layer 3-1) and a normalization layer (LayerNorm 3-1) in sequence, and the other path consists of a linear layer (Linear layer 4-1) and a normalization layer (LayerNorm 4-1) in sequence.

[0016] Input the blood vessel images into the network model and train it according to the following steps:

[0017] Step 1. Use ResBlock1-1 to process the blood vessel image I IN The feature map is obtained through processing.

[0018]

[0019] Step 2. Use the downsampling module 1-1 and the CSVM module 1-1 to... The feature map is obtained through processing. Using the multi-scale edge guidance module EGAA1-1 to The feature map is obtained through processing.

[0020] Step 3. Use the downsampling module 1-2 and the CSVM module 1-2 to... The feature map is obtained through processing. Using the multi-scale edge guidance module EGAA1-2 to The feature map is obtained through processing.

[0021] Step 4. Use the downsampling module 1-3 and the CSVM module 1-3 to... The feature map is obtained through processing. Using the multi-scale edge guidance module EGAA1-3 to The feature map is obtained through processing.

[0022] Step 5. After connecting with The input is processed by the dynamic boundary awareness module (DBA) to obtain the feature map IF. DBA ;

[0023] Step 6. and IF DBA The input is processed by the residual block ResBlock2-1 and then by the upsampling module Upsampling2-1 to obtain the feature map.

[0024] Step 7. and The input is processed by the residual block ResBlock2-2 and then by the upsampling2-2 module to obtain the feature map.

[0025] Step 8. and The input is processed by the residual block ResBlock2-3 and then by the upsampling module Upsampling2-3 to obtain the feature map.

[0026] Step 9. and The input is processed by the residual block ResBlock2-4 to obtain the feature map.

[0027]

[0028] Step 10. Use the Seg Head to process the feature map. Processing yields segmentation results.

[0029] The training process of the CCSA module in CSVM module 1-1, CSVM module 1-2, and CSVM module 1-3 is as follows: Input feature F∈R H×W×C A 1D channel attention map Mc∈R is inferred using the ChannelAttention module. C×1×1 Then, Mc is multiplied by the input feature F, and refined features Fc∈R with channel interest are obtained through the Channel Prior module. C×H×W The refined feature Fc is processed by average pooling layers XAvgPool and YAvgPool respectively to obtain two one-dimensional sequence structures Fc1∈R. C×H Fc2∈R C×WThe two one-dimensional sequence structures are processed by MS-DW Conv, a one-dimensional convolutional layer with kernel sizes of 3, 5, 7, and 9, and two concat function layers. Then, the two features are multiplied element-wise and passed through a group normalization layer (Group Norn) and a ReLU activation function layer.

[0030] The training process of the multi-scale edge guidance modules EGAA1-1, EGAA1-2, and EGAA1-3 is as follows: First, the input feature map is processed by the reverse operation module Reverse1, the Gaussian filter GF, and the deep convolutional layer DWConv. Then, the output result is processed by the convolutional layer Conv1, and then by the convolutional layers Conv2, Conv3, and Conv4 in parallel. The features extracted by the convolutional layers Conv2 and Conv3 are multiplied together and then multiplied by the features extracted by the convolutional layer Conv4. Finally, residual connections are performed with the input image features.

[0031] The training process of the Dynamic Boundary Awareness (DBA) module is as follows: First, the input features are processed by a Dynamic Filter, and then passed through two paths: one path sequentially includes a linear layer (Linear layer 1-1), a normalization layer (LayerNorm 1-1), a convolutional layer (Conv5), an activation function (Sigmoid 1), and a reverse operation module (Reverse 2); the other path sequentially includes a linear layer (Linear layer 2-1), a normalization layer (LayerNorm 2-1), a convolutional layer (Conv6), and an activation function (Sigmoid 2). The results from the two paths are multiplied by the input features respectively, and then the two multiplication results are added together to obtain the feature map. Then Then, two feature maps are obtained by passing the data through two separate paths. and feature map feature map and feature map After connection, it passes through a Softmax activation function layer and is then compared with the feature map. After multiplication, the feature map IF is obtained. DBA .

[0032] This invention relates to a U-shaped encoder-decoder network model for blood vessel image segmentation. The encoder is built around a CSVM module, which is an improved VSS Block structure. Specifically, the DW convolutional layer in the VSS Block is replaced by a CCSA module, thus displaying local and global dependencies in the preserved space in a compressed form, allowing VMamba to access the local and global context before reaching the last token. A multi-scale edge guidance module (EGAA) is placed between the encoder and decoder, using the Laplacian operator to emphasize the boundary features of low-level features for more accurate boundary localization. Furthermore, a dynamic boundary awareness module is placed between the encoder, the multi-scale edge guidance module, and the decoder to aggregate the boundary features of low-level features and the semantic information of high-level features, revealing visual details that are no longer apparent in image segmentation strategies. Experimental results show that this invention surpasses existing state-of-the-art methods, demonstrating its effectiveness and practicality in processing medical images. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the framework of an embodiment of the present invention.

[0034] Figure 2 This is a schematic diagram of the improved VSS Block structure according to an embodiment of the present invention.

[0035] Figure 3 This is a schematic diagram of the CCSA module in an embodiment of the present invention.

[0036] Figure 4 This is a schematic diagram of the structure of the multi-scale edge guidance module EGAA according to an embodiment of the present invention.

[0037] Figure 5 This is a schematic diagram of the structure of the multi-scale edge guidance module EGAA according to an embodiment of the present invention. Detailed Implementation

[0038] This invention discloses a blood vessel image segmentation method based on visual Mamba context-aware semantics. The method uses an encoder-decoder U-shaped network model for blood vessel image segmentation. The encoder is constructed with a CSVM module at its core. The CSVM module is as follows: Figure 2The improved VSS Block structure shown specifically replaces the DW convolutional layer before SS2D in VMamba's VSS Block with a CCSA module, thereby displaying local and global dependencies in the preserved space in a compressed form. A multi-scale edge guidance module is set between the encoder and decoder to emphasize the boundary features of the low-level features using the Laplacian operator. At the same time, a dynamic boundary awareness module is set between the encoder, the multi-scale edge guidance module and the decoder to aggregate the boundary features of the low-level features and the semantic information of the high-level features.

[0039] The specific framework (CSEM-Net) is as follows: Figure 1 As shown:

[0040] The encoder, from top to bottom, consists of a residual block ResBlock1-1 (ResBlock×2), a downsampling module Down sampling1-1, a CSVM module 1-1 (CSVM×2), a downsampling module Down sampling1-2, a CSVM module 1-2 (CSVM×2), a downsampling module Down sampling1-3, and a CSVM module 1-3 (CSVM×2); the CCSA module structure is as follows. Figure 3 As shown: The system consists of a ChannelAttention module, a ChannelPrior module, two average pooling layers XAvgPool and YAvgPool, a one-dimensional convolutional layer MS-DW Conv with kernel sizes of 3, 5, 7 and 9 respectively, two concat function layers, a group normalization layer Group Norn and a ReLU activation function layer.

[0041] The decoder consists of, from bottom to top, a residual block ResBlock2-1 (ResBlock×2), an upsampling module Upsampling2-1, a residual block ResBlock2-2 (ResBlock×2), an upsampling module Upsampling2-2, a residual block ResBlock2-3 (ResBlock×2), an upsampling module Upsampling2-3, a residual block ResBlock2-4 (ResBlock×2), and a Seg Head;

[0042] The residual block ResBlock1-1 and the residual block ResBlock2-4 are connected through the multi-scale edge guidance module EGAA1-1; the CSVM module 1-1 and the upsampling module Upsampling2-3 are connected through the multi-scale edge guidance module EGAA1-2; and the CSVM module 1-2 and the upsampling module Upsampling2-2 are connected through the multi-scale edge guidance module EGAA1-3.

[0043] The CSVM module 1-3 and the residual block ResBlock 2-1 are connected by one jump connection and the other is connected through the dynamic boundary awareness module DBA.

[0044] The multi-scale edge guidance modules EGAA1-1, EGAA1-2, and EGAA1-3 have the same structure, such as... Figure 4 As shown: There are parallel reverse operation modules Reverse1, Gaussian filter GF, and deep convolutional layer DWConv (3×3). The outputs of the reverse operation module Reverse1, Gaussian filter GF and deep convolutional layer DWConv pass through convolutional layer Conv1 (1×1) and are then connected to parallel convolutional layers Conv2 (1×3), Conv3 (3×1) and Conv4 (3×3).

[0045] The dynamic boundary sensing module DBA, as follows: Figure 5 The diagram shows a Dynamic Filter. The output of the Dynamic Filter is divided into two paths: one path consists of a Linear layer 1-1, a normalization layer LayerNorm 1-1, a Conv5 (1×1) convolutional layer, an activation function Sigmoid 1, and a Reverse operation module Reverse 2; the other path consists of a Linear layer 2-1, a normalization layer LayerNorm 2-1, a Conv6 (1×1) convolutional layer, and an activation function Sigmoid 2. The output features are then passed through the two paths and connected to the Softmax activation function layer. One path consists of a Linear layer 3-1 and a normalization layer LayerNorm 3-1; the other path consists of a Linear layer 4-1 and a normalization layer LayerNorm 4-1.

[0046] Input the blood vessel images into the network model and train it according to the following steps:

[0047] Step 1. Use ResBlock1-1 to process the blood vessel image I IN The feature map is obtained through processing.

[0048]

[0049] Step 2. Use the downsampling module 1-1 and the CSVM module 1-1 to... The feature map is obtained through processing. Using the multi-scale edge guidance module EGAA1-1 to The feature map is obtained through processing.

[0050] Step 3. Use the downsampling module 1-2 and the CSVM module 1-2 to... The feature map is obtained through processing. Using the multi-scale edge guidance module EGAA1-2 to The feature map is obtained through processing.

[0051] Step 4. Use the downsampling module 1-3 and the CSVM module 1-3 to... The feature map is obtained through processing. Using the multi-scale edge guidance module EGAA1-3 to The feature map is obtained through processing.

[0052] Step 5. After connecting with The input is processed by the dynamic boundary awareness module (DBA) to obtain the feature map IF. DBA ;

[0053] Step 6. and IF DBA The input is processed by the residual block ResBlock2-1 and then by the upsampling module Upsampling2-1 to obtain the feature map.

[0054] Step 7. and The input is processed by the residual block ResBlock2-2 and then by the upsampling2-2 module to obtain the feature map.

[0055] Step 8. and The input is processed by the residual block ResBlock2-3 and then by the upsampling module Upsampling2-3 to obtain the feature map.

[0056] Step 9. and The input is processed by the residual block ResBlock2-4 to obtain the feature map.

[0057]

[0058] Step 10. Use the Seg Head to process the feature map. Processing yields segmentation result I. Output .

[0059] The training process of the CCSA module in CSVM module 1-1, CSVM module 1-2, and CSVM module 1-3 is as follows: Input feature F∈R H×W×C A 1D channel attention map Mc∈R is inferred using the ChannelAttention module. C×1×1 Then, Mc is multiplied by the input feature F, and refined features Fc∈R with channel interest are obtained through the Channel Prior module. C×H×W The refined feature Fc is processed by average pooling layers XAvgPool and YAvgPool respectively to obtain two one-dimensional sequence structures Fc1∈R. C×H Fc2∈R C×W The two one-dimensional sequence structures are processed by MS-DW Conv, a one-dimensional convolutional layer with kernel sizes of 3, 5, 7, and 9, and two concat function layers. Then, the two features are multiplied element-wise and passed through a group normalization layer (Group Norn) and a ReLU activation function layer.

[0060] The training process of the multi-scale edge guidance modules EGAA1-1, EGAA1-2, and EGAA1-3 is as follows: First, the input feature map is processed by the reverse operation module Reverse1, the Gaussian filter GF, and the deep convolutional layer DWConv. Then, the output result is processed by the convolutional layer Conv1, and then by the convolutional layers Conv2, Conv3, and Conv4 in parallel. The features extracted by the convolutional layers Conv2 and Conv3 are multiplied together and then multiplied by the features extracted by the convolutional layer Conv4. Finally, residual connections are performed with the input image features.

[0061] The training process of the Dynamic Boundary Awareness (DBA) module is as follows: First, the input features are processed by a Dynamic Filter, and then passed through two paths: one path sequentially includes a linear layer (Linear layer 1-1), a normalization layer (LayerNorm 1-1), a convolutional layer (Conv5), an activation function (Sigmoid 1), and a reverse operation module (Reverse 2); the other path sequentially includes a linear layer (Linear layer 2-1), a normalization layer (LayerNorm 2-1), a convolutional layer (Conv6), and an activation function (Sigmoid 2). The results from the two paths are multiplied by the input features respectively, and then the two multiplication results are added together to obtain the feature map. Then Then, two feature maps are obtained by passing the data through two separate paths. and feature map feature map and feature map After connection, it passes through a Softmax activation function layer and is then compared with the feature map. After multiplication, the feature map IF is obtained. DBA .

[0062] experiment:

[0063] I. Algorithm Comparison Experiments on Public Datasets

[0064] The performance results of this invention (CSEM-Net) on the Fives dataset, compared with some of the most relevant models including U-Net, R2U-Net, Attention-UNet, Global Convolutional Network (GCN), Deeplab V3+, Selective Kernel (SK), CBAM, PSPNet, ENet, SegNet, SwinUNet, TransUNet, and sgat-net, are shown in Table 1. The results indicate that CSEM-Net achieves optimal segmentation performance in most cases. First, CSEM-Net outperforms state-of-the-art methods such as SGAT-Net and Skelcon in the IOU metric, improving by 0.20% and 0.35%, respectively. IOU measures the overlap between the predicted segmented region and the ground truth labeled region. The IOU of this invention effectively demonstrates that its method avoids over-segmentation and under-segmentation. Second, the SE metric of this invention achieves a level comparable to methods other than CBAM (93.30%) (91.87%). SE reflects insufficient segmentation of the foreground object, indicating that CSEM-Net can successfully segment more retinal vessels. Furthermore, CSEM-Net achieved near-optimal performance in the SP metric, only 0.10% lower than Skelcon. SP is used to evaluate background oversegmentation, and performance comparisons with existing methods show that CSEM-Net can meet the classification requirements for background pixels. For the comprehensive evaluation metrics of ACC and F1, CSEM-Net outperforms state-of-the-art methods, including Genetic U-Net, Skelcon, and SGAT-Net. The ACC metric considers both foreground and background segmentation results; CSEM-Net achieved a performance of 99.07% on Fives, exceeding state-of-the-art performance by 0.10%, 0.31%, and 0.21%, respectively. The F1 metric can be seen as a trade-off between sensitivity and precision, providing a more balanced reflection of vascular pixel classification accuracy. CSEM-Net achieved an F1 score of 91.11%, exceeding the aforementioned three methods by 0.40%, 0.47%, and 0.60%. These improvements in comprehensive metrics demonstrate that CSEM-Net not only accurately segments retinal vessels but also effectively suppresses background noise. To evaluate the changes in segmentation results under different probability thresholds, the area under the receiver operating characteristic (ROC) curve, i.e., AUC, was also calculated. The CSEM-Net of this invention achieved metrics comparable to Genetic U-Net and Skelcon, but slightly lower than SGAT-Net. The competitive AUC metric indicates that CSEM-Net can more confidently distinguish retinal vessels from the background.

[0065] Table 1

[0066]

[0067] II. Algorithm Comparison Experiments on Private Datasets

[0068] This invention compared the performance of U-Net, UNet++, TransUnet, Swin-UNet, VM-UNet, and the proposed CSEM-Net on coronary artery images. The training strategy involved randomly selecting 40 images for training and the remaining 10 for testing. The comparative experimental results are shown in Table 2. As can be seen from the results in Table 2, the proposed CSEM-Net achieved significant results in the coronary artery segmentation task, achieving the best performance in sensitivity, specificity, accuracy, F1 score, and AUC, and the second-best performance in IOU, approximately 0.25% lower than the best performing SwinUNet.

[0069] Experimental results demonstrate that the CSEM-UNet model of this invention significantly improves the accuracy of medical image segmentation, especially in vascular imaging, through its innovative CSVM module, Dynamic Edge Aggregation (DBA) module, and Multi-Scale Edge Guiding (EGAA) module. The CSVM module combines the linear time complexity advantage of Mamba with the global feature selection capability of CCSA, while the DBA module enhances edge information in the feature map by simulating the biological visual perception process, effectively improving detail reconstruction. Furthermore, the introduction of the EGAA module integrates traditional edge detection with deep learning methods, improving the model's ability to detect weak boundaries. This surpasses existing state-of-the-art methods, demonstrating its effectiveness and practicality in processing medical images.

Claims

1. A blood vessel image segmentation method based on visual Mamba context-aware semantics, which is based on an encoder-decoder U-shaped network model for blood vessel image segmentation, characterized in that: The encoder is built around the CSVM module, which is an improved VSS module structure. Specifically, the DW convolutional layer before SS2D in the VSS module is replaced by a CCSA module, thereby displaying and preserving local and global dependencies in the space in a compressed form. A multi-scale edge guidance module is set between the encoder and the decoder, which uses the Laplacian operator to emphasize the boundary features of the low-level features. At the same time, a dynamic boundary awareness module is set between the encoder, the multi-scale edge guidance module and the decoder to aggregate the boundary features of the low-level features and the semantic information of the high-level features. The encoder, from top to bottom, consists of a residual block ResBlock1-1, a downsampling module Down sampling1-1, a CSVM module 1-1, a downsampling module Down sampling1-2, a CSVM module 1-2, a downsampling module Down sampling1-3, and a CSVM module 1-3; the CCSA consists of a channel attention module Channel Prior, a channel prior module Channel Prior, two average pooling layers XAvgPool and YAvgPool, a one-dimensional convolutional layer MS-DW Conv with kernel sizes of 3, 5, 7, and 9, two concat function layers, a group normalization layer Group Norn, and a ReLU activation function layer; The decoder consists of, from bottom to top, a residual block ResBlock2-1, an upsampling module Up sampling2-1, a residual block ResBlock2-2, an upsampling module Up sampling2-2, a residual block ResBlock2-3, an upsampling module Upsampling2-3, a residual block ResBlock2-4, and a Seg Head; The residual block ResBlock1-1 and the residual block ResBlock2-4 are connected through the multi-scale edge guidance module EGAA1-1; the CSVM module 1-1 and the upsampling module Upsampling2-3 are connected through the multi-scale edge guidance module EGAA1-2; and the CSVM module 1-2 and the upsampling module Upsampling2-2 are connected through the multi-scale edge guidance module EGAA1-3. The CSVM module 1-3 and the residual block ResBlock 2-1 are connected by one skip connection and the other connection is through the dynamic boundary sensing module DBA. The multi-scale edge guidance modules EGAA1-1, EGAA1-2, and EGAA1-3 have the same structure, and are provided with parallel reverse operation modules Reverse1, Gaussian filters GF, and deep convolutional layers DWConv. The outputs of the reverse operation modules Reverse1, Gaussian filters GF, and deep convolutional layers DWConv pass through convolutional layer Conv1 and are then connected to parallel convolutional layers Conv2, Conv3, and Conv4. The Dynamic Boundary Awareness Module (DBA) is equipped with a Dynamic Filter. The output of the Dynamic Filter is divided into two paths: one path consists of a linear layer (Linear layer 1-1), a normalization layer (LayerNorm 1-1), a convolutional layer (Conv 5), an activation function (Sigmoid 1), and a reverse operation module (Reverse 2) in sequence; the other path consists of a linear layer (Linear layer 2-1), a normalization layer (LayerNorm 2-1), a convolutional layer (Conv 6), and an activation function (Sigmoid 2) in sequence. The output features are then passed through two paths and connected to the Softmax activation function layer. One path has a linear layer (Linearlayer3-1) and a normalized layer (Layer Norm3-1) in sequence, and the other path has a linear layer (Linearlayer4-1) and a normalized layer (Layer Norm4-1) in sequence.

2. The blood vessel image segmentation method based on visual Mamba context-aware semantics according to claim 1, characterized in that... Input the blood vessel images into the network model and train it according to the following steps: Step 1. Use ResBlock1-1 to process blood vessel images. The feature map is obtained through processing. ; Step 2. Use the downsampling module 1-1 and the CSVM module 1-1 to... The feature map is obtained through processing. ; Using the multi-scale edge guidance module EGAA1-1 to The feature map is obtained through processing. ; Step 3. Use the downsampling module 1-2 and the CSVM module 1-2 to... The feature map is obtained through processing. ; Using the multi-scale edge guidance module EGAA1-2 to The feature map is obtained through processing. ; Step 4. Use the downsampling module 1-3 and the CSVM module 1-3 to... The feature map is obtained through processing. ; Using the multi-scale edge guidance module EGAA1-3 to The feature map is obtained through processing. ; Step 5. , , After connecting with The input is processed by the dynamic boundary awareness module (DBA) to obtain the feature map. ; Step 6. and The input is processed by the residual block ResBlock2-1 and then by the upsampling module Upsampling2-1 to obtain the feature map. ; Step 7. and The input is processed by the residual block ResBlock2-2 and then by the upsampling2-2 module to obtain the feature map. ; Step 8. and The input is processed by the residual block ResBlock2-3 and then by the upsampling module Upsampling2-3 to obtain the feature map. ; Step 9. and The input is processed by the residual block ResBlock2-4 to obtain the feature map. ; Step 10. Use the Seg Head to process the feature map. Processing yields segmentation results. .

3. The blood vessel image segmentation method based on visual Mamba context-aware semantics according to claim 2, characterized in that... The training process of the CCSA module in CSVM module 1-1, CSVM module 1-2, and CSVM module 1-3 is as follows: Input feature F∈R H×W×C A 1D channel attention map Mc ∈ R is inferred using the channel attention module. C×1×1 Then, Mc is multiplied by the input feature F, and refined features Fc ∈ R with channel interest are obtained through the Channel Prior module. C×H×W The refined feature Fc is processed by average pooling layers XAvgPool and YAvgPool respectively to obtain two one-dimensional sequence structures Fc1 ∈ R. C×H Fc2 ∈ R C×W The two one-dimensional sequence structures are processed by MS-DW Conv, a one-dimensional convolutional layer with kernel sizes of 3, 5, 7, and 9, and two concat function layers. Then, the two features are multiplied element-wise and passed through a group normalization layer (Group Norn) and a ReLU activation function layer.

4. The blood vessel image segmentation method based on visual Mamba context-aware semantics according to claim 3, characterized in that... The training process of the multi-scale edge guidance modules EGAA1-1, EGAA1-2, and EGAA1-3 is as follows: First, the input feature map is processed by the reverse operation module Reverse1, the Gaussian filter GF, and the deep convolutional layer DWConv. Then, the output result is processed by the convolutional layer Conv1, and then by the convolutional layers Conv2, Conv3, and Conv4 in parallel. The features extracted by the convolutional layers Conv2 and Conv3 are multiplied together and then multiplied by the features extracted by the convolutional layer Conv4. Finally, residual connections are performed with the input image features.

5. The blood vessel image segmentation method based on visual Mamba context-aware semantics according to claim 4, characterized in that... The training process of the Dynamic Boundary Awareness (DBA) module is as follows: First, the input features are processed by a Dynamic Filter, and then passed through two paths: one path sequentially includes a linear layer (Linear layer 1-1), a normalization layer (Layer Norm 1-1), a convolutional layer (Conv 5), an activation function (Sigmoid 1), and a reverse operation module (Reverse 2); the other path sequentially includes a linear layer (Linear layer 2-1), a normalization layer (Layer Norm 2-1), a convolutional layer (Conv 6), and an activation function (Sigmoid 2). The results from the two paths are multiplied by the input features respectively, and then the two multiplication results are added together to obtain the feature map. ;Then Then, two feature maps are obtained by passing the data through two separate paths. and feature map , feature map and feature map After connection, it passes through a Softmax activation function layer and is then compared with the feature map. The feature map is obtained after multiplication. .

Images