A blood vessel segmentation method, device and medium based on U-Net fusion of multi-scale dilated convolution

Abstract

This invention relates to a method, device, and medium for blood vessel segmentation based on U-Net fusion with multi-scale dilated convolution. The method involves inputting a retinal image to be segmented into a trained DRU-Net model to obtain the blood vessel segmentation result. The DRU-Net model uses U-Net as a baseline, replacing traditional convolution with multi-scale dilated residual modules in the encoder and decoder parts, and embedding multi-level detail attention modules at skip connections. The multi-scale dilated residual module includes multiple parallel heterogeneous dilated convolution paths with different dilation rates, aggregating multi-scale features through multi-path convolution. The multi-level detail attention module aggregates information in different spatial dimensions through efficient multi-scale attention. Efficient local attention is employed, utilizing strip pooling to extract features in the horizontal and vertical directions respectively. Compared with existing technologies, this invention achieves robust modeling of blood vessel morphology from local details to global distribution in complex backgrounds, significantly reducing the breakage and blurring of small blood vessels.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing, and in particular to a blood vessel segmentation method based on U-Net fusion of multi-scale dilated convolution. Background Technology

[0002] In the field of medical image segmentation, the following challenges exist for retinal vessel segmentation.

[0003] 1) The topological continuity of small blood vessels is easily broken: The encoder of traditional U-Net relies on local convolution operation, and the receptive field is limited to adjacent pixels. During the downsampling process, the blood vessel branch information will be lost, and the skip connections will be interfered with by noise. It is difficult to capture the overall topological structure of blood vessels, resulting in blurred or broken blood vessel edges.

[0004] 2) Insufficient adaptation to multi-scale features: The main arteries of the retinal vessels are typically larger than 100 μ m and capillaries are less than 10 μ m, the diameter of blood vessels can span tens of times, and it is difficult to take into account the detailed features of the global network and local terminals with a fixed convolution kernel size.

[0005] 3) Dependence on labeled data: Supervised learning relies on a large amount of labeled data, and datasets need to be manually labeled by experts. However, medical image labeling is costly and easily affected by subjectivity.

[0006] U-Net, a landmark work, was the first to adopt a symmetrical encoder-decoder structure and fused shallow detail features with deep semantic information through skip connections, laying a solid foundation for blood vessel segmentation. Current researchers have made various improvements to the U-Net framework. To enhance the feature extraction capability for microvessels, Jha et al. proposed ResUNet++, which significantly improved the recognition performance of capillary branches by introducing residual connections and channel attention mechanisms. Duan et al. introduced pre-trained deformable convolutional layers into the encoder to enhance adaptability to local morphology. Wang et al. introduced a CHFM module for hierarchical multimodal feature fusion, combined with high-level semantic weights to reduce modal differences. However, their algorithm is still affected by capillary interference and lesion areas, and its generalization ability on severe pathological images such as massive hemorrhage and edema has not been fully validated. Ma et al. proposed a dynamic directional attention module used at skip connections to model horizontal and vertical attention separately, enhancing the perception of blood vessel orientation and thus better preserving small blood vessel branches and complex bifurcation structures. However, the authors also pointed out that model performance may degrade in extreme cases of extremely high noise or extremely high blood vessel curvature. This means that its generalization ability in complex pathological scenarios still has room for improvement. Summary of the Invention

[0007] The purpose of this invention is to solve the two major problems in retinal vessel segmentation: the easy breakage of vessel topological continuity and the insufficient adaptability of multi-scale features. It provides a retinal vessel segmentation method based on U-Net fusion of multi-scale dilated convolution.

[0008] The objective of this invention can be achieved through the following technical solutions: As a first aspect of the present invention, a blood vessel segmentation method based on U-Net fusion of multi-scale dilated convolution is provided, wherein the method inputs the retinal image to be segmented into a trained DRU-Net model to obtain the blood vessel segmentation result; The DRU-Net model uses U-Net as a baseline, and adopts multi-scale dilated residual modules in the encoder and decoder to replace traditional convolution, and embeds multi-level detail attention modules at skip connections. The multi-scale dilated residual module includes multiple parallel heterogeneous dilated convolution paths with different dilation rates, which aggregate multi-scale features through multi-path convolution. The multi-level detail attention module aggregates information in different spatial dimensions through efficient multi-scale attention; it employs efficient local attention and utilizes strip pooling to extract features in the horizontal and vertical directions respectively.

[0009] As a preferred technical solution, the multi-scale dilation residual module is specifically as follows: The input feature map is decoupled into four sets of parallel feature representations; The three sets of features extract multi-scale contextual information through three parallel branches, and each branch achieves feature transformation through 3×3 dilated convolution with different dilation rates; the fourth set of features represents the spatial attention features extracted by the convolutional block attention module. The convolutional features obtained from the three parallel branches are added element-wise and batch normalization is applied. Then, the SE module is used to enhance the feature representation of the channel dimension, and the enhanced feature representation of the channel dimension is fused with the spatial attention features element-wise. The element-wise fused features are transformed using 3×3 convolution and ReLU activation function, and aggregated into unified enhanced features.

[0010] As a preferred technical solution, the first branch of the three parallel branches includes two dilated convolutional layers with a dilation rate of 1 and a kernel size of 3×3, which are connected in sequence and have batch normalization.

[0011] As a preferred technical solution, the second branch of the three parallel branches includes a dilated convolutional layer with an expansion rate of 1, a kernel size of 3×3 and batch normalization, and two dilated convolutional layers with an expansion rate of 3, a kernel size of 3×3 and batch normalization, connected in sequence.

[0012] As a preferred technical solution, the third branch of the three parallel branches includes two dilated convolutional layers with a dilation rate of 3 and a kernel size of 3×3 with batch normalization, connected in sequence, and a dilated convolutional layer with a dilation rate of 5 and a kernel size of 3×3 with batch normalization.

[0013] As a preferred technical solution, the multi-level detail attention module includes efficient multi-scale attention and efficient local attention; it performs spatial pooling in the horizontal and vertical directions through efficient multi-scale attention and strip pooling through efficient local attention; and it performs residual fusion between the output of efficient multi-scale attention and efficient local attention and the input feature map before outputting.

[0014] As a preferred technical solution, the efficient multi-scale attention divides the input feature map into three-dimensional sub-feature maps based on the number of channels and inputs them into two parallel sub-networks respectively. The first sub-network extracts one-dimensional spatial information by performing spatial pooling in the horizontal and vertical directions for each dimension; the one-dimensional spatial information in the two directions is concatenated and fused by a 1×1 convolution, and then a sigmoid activation function is used to generate spatial attention weight values; the spatial attention weight values ​​are then residually connected to the sub-feature maps and group normalization is performed. The second sub-network uses 3×3 convolutions on the sub-feature maps to capture local context and multi-scale structure; The outputs of the two sub-networks are processed by Softmax normalization and average pooling respectively; matrix multiplication is performed on the Softmax normalized result of the first sub-network output and the average pooling result of the second sub-network output; matrix multiplication is performed on the average pooling result of the first sub-network output and the Softmax normalized result of the second sub-network output. The results of the two matrix multiplications are concatenated and residually connected with the sub-feature maps to output enhanced multi-scale features.

[0015] As a preferred technical solution, the efficient local attention uses strip pooling to extract features in the horizontal and vertical directions respectively. The feature vectors in the horizontal and vertical directions are processed by group normalization and non-linear activation function Sigmoid to generate position attention predictions in the two directions. The predicted position attentions in the two directions are multiplied element-wise to obtain the position attention.

[0016] As a second aspect of the present invention, an electronic device is provided, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the blood vessel segmentation method based on U-Net fusion multi-scale dilated convolution as described above.

[0017] As a third aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the blood vessel segmentation method based on U-Net fusion multi-scale dilated convolution as described above.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1) Based on U-Net, this invention proposes a novel DRU-Net architecture that integrates solutions to the challenges of vascular continuity and multi-scale features. By systematically integrating the Multi-Scale Dilated Residual Module (MDRM) with the Multi-Level Detail Attention Module (MLDA), a synergistic enhancement of vascular topological continuity and multi-scale feature adaptability is achieved.

[0019] 2) The present invention embeds a multi-scale dilated residual module (MDRM) in the encoder and decoder. This module combines multi-scale dilated convolution with residual connection to expand the receptive field of the model, enabling the model to more comprehensively understand and utilize local details and global context.

[0020] 3) This invention designs an attention fusion mechanism inspired by the anatomical characteristics of blood vessels. The proposed MLDA module performs structured modeling of the spatial topological relationship of blood vessels based on the specific anatomical distribution patterns of retinal vessels along the horizontal and vertical directions, and by simulating the tracking and analysis logic in clinical image reading. Attached Figure Description

[0021] Figure 1 This is a flowchart of a blood vessel segmentation method based on U-Net fusion with multi-scale dilated convolution according to the present invention.

[0022] Figure 2 This is a schematic diagram of the architecture of the DRU-Net model proposed in this invention.

[0023] Figure 3 This is a schematic diagram of the MDRM module structure proposed in this invention.

[0024] Figure 4 This is a schematic diagram of the MLDA module structure proposed in this invention.

[0025] Figure 5The following are schematic diagrams of the retinal vessel segmentation process in a specific embodiment of the present invention: a) original complete retinal image; b) complete real label; c) partial schematic diagram of image block a; d) partial schematic diagram of image block b; e) partial schematic diagram of image block c; f) partial schematic diagram of image block d. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0027] Example 1 This invention proposes a retinal vessel segmentation method based on U-Net fusion with multi-scale dilated convolution. To achieve accurate segmentation of retinal vessels, a novel DRU-Net architecture is proposed to address the challenges of vessel continuity and multi-scale feature integration. By systematically integrating the Multi-Scale Diverse Reception Module (MDRM) with the Multi-Level Detail Attention Module (MLDA), a synergistic enhancement of vessel topological continuity and multi-scale feature adaptability is achieved, thereby improving segmentation accuracy.

[0028] Step S1: Data preparation.

[0029] Data collection: Collect a large amount of labeled training data, including input images and corresponding segmentation masks.

[0030] Data preprocessing: The image is normalized to ensure that pixel values ​​are within a certain range (e.g., 0 to 1). For segmentation masks, they are usually converted into binary images (foreground is 1, background is 0).

[0031] Data augmentation: Increase the diversity of data through data augmentation operations such as random horizontal flipping, vertical flipping, and rotation of [90, 180, 270] degrees.

[0032] Step S2: DRU-Net model construction.

[0033] Encoder-decoder architecture: Constructing the improved U-Net model as follows Figure 1 As shown, the framework uses U-Net as the baseline network structure. In the encoder and decoder, the Multi-Scale Dilated Residual Module (MDRM) aggregates multi-scale features through multi-path convolution to enhance feature representation capabilities. The Multi-Level Detail Attention Module (MLDA) integrates the multi-directional spatial pooling mechanism of EMA and the strip pooling mechanism of ELA to optimize detail perception and suppress noise.

[0034] The structured design of the Multi-Scale Dilated Residual Module (MDRM) replaces traditional convolutional methods. The MDRM module simulates multi-level feature extraction through parallel dilated convolutional paths (dilation rates 1, 3, and 5), and preserves local details by incorporating residual connections. This design is inspired by retinal vascular anatomy (the correlation between local microvascular leakage, mesoscale tissue edema, and global structural remodeling), rather than simply stacking convolutional layers. Simultaneously, the CBAM attention module enhances and suppresses noise in the original features, combining them with SE-enhanced features through additive operations. Robust local-to-global modeling is achieved through formulaic concatenation and feature enhancement.

[0035] MDRM module architecture as follows Figure 2 As shown, multiple convolutional layers with different dilation rates are integrated to supplement and explore multi-scale feature information. Specifically, the feature map output by the encoder or decoder is decoupled into four sets of parallel feature representations. Each set of representations is then processed through heterogeneous feature extraction paths and finally aggregated into a unified enhanced feature.

[0036] The input image extracts multi-scale contextual information through three parallel branches, each branch performing feature transformation via 3×3 dilated convolutions with varying dilation rates. Then, a BN+SE module enhances the channel-dimensional feature representation and fuses it element-wise with the spatial attention features output by CBAM. Finally, feature transformation is completed using 3×3 convolutions and the ReLU activation function, achieving robust extraction and efficient representation of complex features.

[0037] First, dilated convolutional layers with different dilation rates and 3×3 kernel size, along with batch normalization, are used to mine features. After each convolutional operation, the layers are concatenated to supplement fine-grained features and explore small target features. This process can be expressed by the following formula: After obtaining the convolutional features from the three branches above, batch normalization is applied, and then the SE module is used to enhance valuable features and suppress irrelevant features.

[0038] The fourth branch uses a Convolutional Block Attention Module (CBAM) to process the input image, enhancing the original features output by the encoder and suppressing noise.

[0039] Finally, by combining the element-wise addition operation with the results of the first three SE operations and the output features of the fourth branch, a convolutional layer with a kernel size of 3×3 is used for feature extraction and denoising. Then, the ReLU activation function is used to complete the feature transformation. This process can be summarized by the following formula: in, It is the result of four parallel feature mappings of the MDRM module; It's a splicing operation; This represents the multi-channel characteristics of the encoder output; This represents a dilated convolutional layer with batch normalization. This refers to the dilation rate.

[0040] In clinical diagnosis, single-scale features are insufficient for a comprehensive assessment of lesions: microvascular leakage (local), surrounding tissue edema (mesoscale), and global retinal structural remodeling (global) are interconnected pathological processes. To address this issue, this invention introduces a Multi-Scale Dilated Residual Module (MDRM) in the encoder and decoder sections. This module integrates multi-scale dilated convolution and residual connections to resolve the medical challenge of the disconnect between local lesions and the global context in retinal vascular lesion assessment. MDRM simulates this multi-level assessment logic through three parallel branches (dilation rates of 1, 3, and 5): the small dilation rate branch focuses on subtle abnormalities such as local microvascular leakage, the medium dilation rate branch senses the diffusion effect of lesions on the surrounding nerve fiber layer, and the large dilation rate branch captures global structural changes in key areas such as the optic disc or macula. Simultaneously, a CBAM module is used to enhance and suppress noise in the original features. Residual connections further simulate the bypass compensation mechanism of retinal vessels, ensuring that local features are not lost in the deep network, thereby achieving robust modeling of vascular morphology from local details to global distribution in complex contexts.

[0041] The structured design of the Multi-Level Detail Attention (MLDA) module involves using MLDA modules at skip connections. The MLDA module includes multi-directional spatial pooling (horizontal / vertical) for Efficient Multi-Scale Attention (EMA) and strip pooling for Efficient Local Attention (ELA). The EMA module, through horizontal and vertical spatial pooling, enhances the perception of horizontal context regarding the orientation of the main vascular vessels and vertical context regarding the hierarchical distribution of capillaries, respectively, simulating the analytical logic of tracing along the main vascular axis and cross-layer comparison in clinical image interpretation. ELA, through strip pooling, further enhances the perception of the continuity of small vessels locally, suppressing interference from noise such as exudates. This bidirectional attention mechanism is essentially a structured modeling of the spatial topological relationship of retinal vessels, ensuring the preservation of anatomical continuity of vessels during feature fusion, thereby significantly reducing the breakage and blurring of small vessels.

[0042] MLDA module architecture as follows Figure 3 As shown, the MLDA module combines the efficient multi-scale attention mechanism (EMA) with the efficient local attention mechanism (ELA) using strip pooling, both prioritizing computational efficiency. EMA focuses on multi-scale fusion, while ELA emphasizes local enhancement, aiming to improve the model's ability to perceive multi-scale or local features. This addresses the issue of retinal vascular topological continuity being susceptible to noise interference in skip connections.

[0043] EMA divides the input feature map into three sub-feature maps based on the number of channels, and inputs each sub-feature map into two parallel sub-networks. The first sub-network performs spatial pooling in both the horizontal and vertical directions for each dimension to extract one-dimensional spatial information. The two one-dimensional spatial information are concatenated and fused using a 1×1 convolution, then activated by a sigmoid activation function to generate spatial attention weights. These spatial attention weights are then residual-connected to the sub-feature maps and grouped normalized. The second sub-network uses 3×3 convolutions on the sub-feature maps to capture local context and multi-scale structures. The outputs of both sub-networks are processed using Softmax normalization and average pooling, respectively. Matrix multiplication is performed between the Softmax normalized result of the first sub-network and the average pooling result of the second sub-network; matrix multiplication is also performed between the average pooling result of the first sub-network and the Softmax normalized result of the second sub-network; the results of the two matrix multiplications are concatenated and residual-connected to the sub-feature maps to output enhanced multi-scale features.

[0044] ELA learns from information at different levels, enabling shallow and deep features to interact and capture richer feature representations. It utilizes one-dimensional convolution to achieve local interaction between two feature vectors, the coverage of which can be adjusted by changing the kernel size. Spatially, strip pooling is used to extract features in the horizontal and vertical directions respectively. The resulting feature vectors are then processed by group normalization and the sigmoid activation function to generate positional attention predictions in both directions. The final positional attention is obtained by multiplying the positional attention in both directions, calculated as follows: In the formula, X Represents the original input feature map. and It refers to positional attention in two directions, where ⊙ represents element-wise multiplication, and... This indicates that the attention maps in two directions are multiplied element-wise.

[0045] Finally, the MLDA module encoder adds its input to the outputs of the EMA and ELA modules, concatenates it with the upsampled feature map of the corresponding layer, and inputs it into the decoder of the corresponding layer.

[0046] Step S3, Training Loop.

[0047] Loss function: Choose an appropriate loss function, such as Dice loss.

[0048] Optimizer: Select the Adam optimizer and set hyperparameters such as the learning rate.

[0049] Initialization: Initialize the model parameters, typically using a Gaussian distribution to initialize the weights to ensure that each feature map in the network has approximately unit variance.

[0050] Forward propagation: The input image is propagated forward through the model to obtain the predicted segmentation mask.

[0051] Loss calculation: Calculate the loss based on the prediction results and the actual segmentation mask.

[0052] Backpropagation: The gradient is calculated through backpropagation, and the model parameters are updated using the optimizer.

[0053] Monitor the training process: record training loss and validation loss, and save model weights periodically.

[0054] Step S4: Verification.

[0055] Validation set: Evaluate model performance on the validation set, calculate validation loss and evaluation metrics (such as the Dice coefficient).

[0056] Adjust hyperparameters: Adjust hyperparameters such as learning rate and batch size based on validation results to improve model performance.

[0057] Step S5: Save the model.

[0058] Save the best model: The weights of the model that performs best on the validation set are saved for subsequent inference or further training.

[0059] Existing methods often focus on single-scale attention, while this invention utilizes MDRM to synergistically optimize multi-scale dilated convolution and residual connections, explicitly modeling the wide range of blood vessel diameters (from >100μm in main arteries to <10μm in capillaries), thus solving the problem of a single receptive field in U-Net. Simultaneously, MLDA, through residual fusion of EMA and ELA, suppresses noise by structurally modeling the spatial topology of blood vessels, rather than relying on a general attention mechanism, significantly reducing vascular rupture. This modular design based on the logic of clinical pathological evolution is unprecedented in existing technologies.

[0060] To verify the segmentation performance of the model, accuracy (ACC), sensitivity (SE), specificity (SP), and F1 score were used as evaluation metrics. Table 1 shows the advantages of the improved U-Net model proposed in this invention in retinal vessel segmentation. This method was compared with existing image segmentation networks, and the results are shown in Table 1.

[0061] Table 1. Model Comparison Results The segmentation results were compared with existing image segmentation networks on the DRIVE dataset. The accuracy, sensitivity, specificity, and F1 score of the proposed DRU-Net method were 0.9651, 0.8122, 0.9856, and 0.8276, respectively. Compared with some current segmentation models, DRU-Net achieved the highest accuracy, specificity, and F1 score, only slightly lower than SCUNet's 0.8134 in sensitivity (0.8122). Compared with LinkNet, which has a higher F1 score, DRU-Net improved accuracy, sensitivity, and specificity by 0.0075, 0.0212, and 0.0050, respectively. These results demonstrate that the proposed DRU-Net performs well on the DRIVE dataset. Figure 5 As shown, the original U-Net model produces blurred and fragmented blood vessels during segmentation. In contrast, this invention displays more complete blood vessels even when segmenting the fine edges of small vessels.

[0062] Example 2 As a second aspect of the present invention, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and, when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the vessel segmentation method based on U-Net fusion multi-scale dilated convolution as described above. In addition to the processors, memory, and interfaces described above, any data processing device in the embodiments may also include other hardware depending on the actual function of the data processing device, which will not be elaborated further.

[0063] Example 3 As a third aspect of the present invention, this application also provides a computer-readable storage medium storing computer instructions thereon, which, when executed by a processor, implement the vessel segmentation method based on U-Net fusion multi-scale dilated convolution as described above. The computer-readable storage medium can be an internal storage unit of any data processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0064] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A blood vessel segmentation method based on U-Net fusion of multi-scale dilated convolution, characterized in that, The method inputs the retinal image to be segmented into a trained DRU-Net model to obtain the blood vessel segmentation result; The DRU-Net model uses U-Net as a baseline, and adopts multi-scale dilated residual modules in the encoder and decoder to replace traditional convolution, and embeds multi-level detail attention modules at skip connections. The multi-scale dilated residual module includes multiple parallel heterogeneous dilated convolution paths with different dilation rates, which aggregate multi-scale features through multi-path convolution. The multi-level detail attention module aggregates information in different spatial dimensions through efficient multi-scale attention; it employs efficient local attention and utilizes strip pooling to extract features in the horizontal and vertical directions respectively.

2. The blood vessel segmentation method based on U-Net fusion of multi-scale dilated convolution as described in claim 1, characterized in that, The multi-scale dilation residual module is specifically as follows: The input feature map is decoupled into four sets of parallel feature representations; The three sets of features extract multi-scale contextual information through three parallel branches, and each branch achieves feature transformation through 3×3 dilated convolution with different dilation rates; the fourth set of features represents the spatial attention features extracted by the convolutional block attention module. The convolutional features obtained from the three parallel branches are added element-wise and batch normalization is applied. Then, the SE module is used to enhance the feature representation of the channel dimension, and the enhanced feature representation of the channel dimension is fused with the spatial attention features element-wise. The element-wise fused features are transformed using 3×3 convolution and ReLU activation function, and aggregated into unified enhanced features.

3. The blood vessel segmentation method based on U-Net fusion of multi-scale dilated convolution as described in claim 2, characterized in that, The first branch of the three parallel branches includes two dilated convolutional layers with a dilation rate of 1 and a kernel size of 3×3, which are connected in sequence and have batch normalization.

4. The blood vessel segmentation method based on U-Net fusion of multi-scale dilated convolution as described in claim 2, characterized in that, The second branch of the three parallel branches includes a dilated convolutional layer with a dilation rate of 1, a kernel size of 3×3, and batch normalization, connected in sequence, and two dilated convolutional layers with a dilation rate of 3, a kernel size of 3×3, and batch normalization.

5. The blood vessel segmentation method based on U-Net fusion of multi-scale dilated convolution according to claim 2, characterized in that, The third branch of the three parallel branches includes two dilated convolutional layers with a dilation rate of 3 and a kernel size of 3×3 with batch normalization, connected in sequence, and a dilated convolutional layer with a dilation rate of 5 and a kernel size of 3×3 with batch normalization.

6. The blood vessel segmentation method based on U-Net fusion of multi-scale dilated convolution as described in claim 1, characterized in that, The multi-level detail attention module includes efficient multi-scale attention and efficient local attention; it performs spatial pooling in the horizontal and vertical directions through efficient multi-scale attention, and strip pooling through efficient local attention. The output of efficient multi-scale attention and efficient local attention is residually fused with the input feature map before output.

7. The blood vessel segmentation method based on U-Net fusion of multi-scale dilated convolution as described in claim 6, characterized in that, The efficient multi-scale attention divides the input feature map into three-dimensional sub-feature maps based on the number of channels and inputs them into two parallel sub-networks respectively. The first sub-network extracts one-dimensional spatial information by performing spatial pooling in the horizontal and vertical directions for each dimension; the one-dimensional spatial information in the two directions is concatenated and fused by a 1×1 convolution, and then a sigmoid activation function is used to generate spatial attention weight values; the spatial attention weight values ​​are then residually connected to the sub-feature maps and group normalization is performed. The second sub-network uses 3×3 convolutions on the sub-feature maps to capture local context and multi-scale structure; The outputs of the two sub-networks are processed by Softmax normalization and average pooling respectively; matrix multiplication is performed on the Softmax normalized result of the first sub-network output and the average pooling result of the second sub-network output; matrix multiplication is performed on the average pooling result of the first sub-network output and the Softmax normalized result of the second sub-network output. The results of the two matrix multiplications are concatenated and residually connected with the sub-feature maps to output enhanced multi-scale features.

8. The blood vessel segmentation method based on U-Net fusion of multi-scale dilated convolution according to claim 6, characterized in that, The efficient local attention uses strip pooling to extract features in the horizontal and vertical directions respectively. The feature vectors in the horizontal and vertical directions are then processed by group normalization and the non-linear activation function Sigmoid to generate position attention predictions in the two directions. The predicted position attentions in the two directions are multiplied element-wise to obtain the position attention.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the blood vessel segmentation method based on U-Net fusion multi-scale dilated convolution as described in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the blood vessel segmentation method based on U-Net fusion multi-scale dilated convolution as described in any one of claims 1-8.

Images