A blood vessel image segmentation method based on deep mutual learning

Abstract

This invention discloses a deep mutual learning-based blood vessel image segmentation method, belonging to the field of medical image segmentation technology. The method includes: a dual-branch mutual learning segmentation model comprising a parallel Attention_Unet_EfficientV2 model and a Base_Unet_EfficientNet model; the deep mutual learning-based dual-branch mutual learning segmentation model performs feature alignment and error correction between the Attention_Unet_EfficientV2 model and the Base_Unet_EfficientNet model through collaborative training; the Attention_Unet_EfficientV2 model is constructed by incorporating an attention mechanism on the basis of the traditional U-Net model; the Base_Unet_EfficientNet model is an improvement on the traditional U-Net model and EfficientNet-B5; the deep mutual learning-based dual-branch mutual learning segmentation model is used to segment the blood vessel image to be segmented, obtaining the blood vessel contours in the image. This method improves the segmentation accuracy of blood vessel images.

Description

Technical Field

[0001] This application relates to the field of medical image segmentation technology, and in particular to a blood vessel image segmentation method based on deep mutual learning. Background Technology

[0002] Cardiovascular diseases and retinal diseases are major threats to human health. Early detection and precise treatment are crucial for improving patients' quality of life. Therefore, efficient and accurate segmentation of these two types of vascular images has significant clinical implications.

[0003] With the development of deep learning, it has made revolutionary progress in the field of medical image analysis. However, existing deep learning models are prone to problems such as blurred boundaries, broken branches, and discontinuous vascular structures when identifying small, low-contrast blood vessels, resulting in low accuracy of vascular image segmentation. Summary of the Invention

[0004] Therefore, it is necessary to provide a blood vessel image segmentation method based on deep mutual learning to address the aforementioned technical problems. This method improves the segmentation accuracy of blood vessel images.

[0005] The following technical solution is adopted in this specification: This specification provides a blood vessel image segmentation method based on deep mutual learning, including: Acquire the image of the blood vessel to be segmented; A two-branch mutual learning segmentation model is obtained; the two-branch mutual learning segmentation model includes a parallel Attention_Unet_EfficientNetV2 model and a Base_Unet_EfficientNet model; the two-branch mutual learning segmentation model performs feature alignment and error correction between the Attention_Unet_EfficientNetV2 model and the Base_Unet_EfficientNet model through co-training; the Attention_Unet_EfficientNetV2 model is built on the basis of the traditional U-Net model by incorporating an attention mechanism; the encoder of the Base_Unet_EfficientNet model is the encoder of EfficientNet-B5, and the decoder is the decoder of the traditional U-Net model. Feature transfer is achieved between the encoder and decoder through skip connections; A bi-branch mutual learning segmentation model is used to segment blood vessel images to obtain the blood vessel contours in the images to be segmented.

[0006] Optionally, the Attention_Unet_EfficientNetV2 model includes an encoder and a decoder; the encoder and decoder achieve feature transfer through cross-layer skip connections, and each skip connection integrates an attention gating module; the attention gating module reconstructs the spatial and channel weights of the shallow features transferred by the encoder through adaptive average pooling, grouped convolution and feature fusion. The encoder of the Attention_Unet_EfficientNetV2 model consists of multiple feature extraction layers connected in series; each feature extraction layer includes two sets of convolutional layers, batch normalization layers, and linear activation layers connected in series; each set of convolutional layers, batch normalization layers, and linear activation layers in the encoder includes a downsampling module between them. The decoder of the Attention_Unet_EfficientNetV2 model consists of multiple feature decoding layers connected in series and a 1×1 convolutional layer; each feature decoding layer includes multiple sets of convolutional layers, batch normalization layers and linear activation layers connected in series; each of the concatenated feature decoding layers in the decoder includes an upsampling module. The Base_Unet_EfficientNet model includes an encoder and a decoder; The Base_Unet_EfficientNet model achieves feature transfer between the encoder and decoder through cross-layer skip connections; The encoder of the Base_Unet_EfficientNet model consists of multiple feature extraction layers connected in series; each feature extraction layer includes two sets of convolutional layers, batch normalization layers, and linear activation layers connected in series; each set of convolutional layers, batch normalization layers, and linear activation layers in the encoder includes a downsampling module between them; The decoder of the Base_Unet_EfficientNet model consists of multiple feature decoding layers connected in series and a 1×1 convolutional layer; each feature decoding layer includes multiple sets of convolutional layers, batch normalization layers and linear activation layers connected in series; each of the concatenated feature decoding layers in the decoder includes an upsampling module.

[0007] Optionally, the Attention_Unet_EfficientNetV2 model also includes a classification mapping convolutional layer and a bilinear interpolation recovery module; the Base_Unet_EfficientNet model also includes a classification mapping convolutional layer and a scale alignment module; the blood vessel image to be segmented is segmented using a two-branch mutual learning segmentation model to obtain the blood vessel contour in the image to be segmented, specifically including: The blood vessel image to be segmented is input into the Attention_Unet_EfficientNetV2 model, and after multiple feature extraction layers connected in series in the encoder, it is downsampled at multiple levels to obtain the first multi-scale encoded features; The first multi-scale encoded feature is fused and decoded by the skip connection and upsampling module guided by the attention gating module to obtain the first decoded output feature Z1. The first decoded output feature Z1 is input into the classification mapping convolutional layer for channel dimension compression processing to obtain the first branch prediction probability map P1. The first branch prediction probability map P1 is then processed by the bilinear interpolation recovery module to obtain the first branch processing result with the same size as the blood vessel image to be segmented. The blood vessel image to be segmented is input into the Base_Unet_EfficientNet model, and the second multi-scale encoded features of the blood vessel image to be segmented are extracted through multiple feature extraction layers of the Base_Unet_EfficientNet model. By using a skip connection in the form of direct splicing, an upsampling module, and a multi-layer feature decoding layer, the second multi-scale encoded feature is restored and decoded to obtain the second decoded output feature Z2. The second decoded output feature Z2 is input into the classification mapping convolutional layer for class probability mapping to obtain the second branch prediction probability map P2. The second branch prediction probability map P2 is input into the scale alignment module for global size restoration processing to obtain the second branch processing result; The results of the first branch processing and the second branch processing are fused by bi-branch inference to obtain the blood vessel contour in the blood vessel image to be segmented.

[0008] Optionally, the encoder of the Attention_Unet_EfficientNetV2 model includes a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, a fourth feature extraction layer, and a fifth feature extraction layer. After multi-level downsampling through multiple feature extraction layers sequentially connected in the encoder, the first multi-scale encoded features are obtained, specifically including: The 512×512 pixel blood vessel image to be segmented is input into the first feature extraction layer of the encoder. After passing through two sets of convolutional layers, batch normalization layers and linear activation layers in sequence, a first feature map with a dimension of 256×256 and a channel number of 24 is obtained. After downsampling the first feature map, it is passed through the second feature extraction layer to obtain a second feature map with a dimension of 128×128 and a number of channels of 48. After downsampling the second feature map, it is passed through the third feature extraction layer to obtain a third feature map with a dimension of 64×64 and a number of channels of 64. After downsampling the third feature map, it is passed through the fourth feature extraction layer to obtain a fourth feature map with a dimension of 32×32 and a number of channels of 160. The fourth feature map is passed through the fifth feature extraction layer to obtain the first multi-scale coding feature with a dimension of 16×16 and a channel number of 1280.

[0009] Optionally, the first multi-scale encoded features are fused and decoded through a skip connection and upsampling module guided by an attention gating module to obtain the first decoded output feature Z1, specifically including: The multi-scale encoded features with dimensions of 16×16 and 1280 channels are upsampled to a dimension of 32×32 and then concatenated with the attention-weighted fourth feature map in terms of channel dimension. After passing through the first decoding layer in the decoder, the first decoded feature with 160 channels is obtained. The first decoded feature is upsampled and then concatenated with the attention-weighted third feature map in the channel dimension. After passing through the second decoding layer in the decoder, a second decoded feature with 64 channels is obtained. The second decoded feature is upsampled and then concatenated with the attention-weighted second feature map in the channel dimension. After passing through the third decoding layer in the decoder, a third decoded feature with 48 channels is obtained. The third decoded feature is upsampled and then concatenated with the attention-weighted first feature map in the channel dimension. After passing through the fourth decoding layer in the decoder, a fourth decoded feature with 24 channels is obtained. The fourth decoded feature is passed through a 1×1 convolutional layer to obtain the first decoded output feature Z1.

[0010] Optionally, the training process of the two-branch mutual learning segmentation model specifically includes: Acquire sample blood vessel images and annotate the blood vessel contours in the sample blood vessel images to obtain the training set; In the first training phase, the Attention_Unet_EfficientNetV2 model and the Base_Unet_EfficientNet model in the dual-branch mutual learning segmentation model are trained based on the training set until the hard loss values ​​of the Attention_Unet_EfficientNetV2 model and the Base_Unet_EfficientNet model reach their minimum. In the second training phase, based on the training set, soft loss generated by knowledge distillation is introduced, and a cosine growth strategy is used to gradually increase the weight coefficient of soft loss in the total loss. The two-branch mutual learning segmentation model is trained until the two-branch mutual learning segmentation model smoothly transitions from independent training to the co-distillation state. In the third training phase, based on the training set, the weights between hard loss and soft loss are fixed and the temperature parameter is gradually reduced as the training rounds increase. The prediction probability distribution is gradually sharpened to force the bi-branch mutual learning segmentation model to focus on the key anatomical features of blood vessels. In the fourth training phase, based on the training set, the parameters of the Attention_Unet_EfficientNetV2 model and the Base_Unet_EfficientNet model are adjusted until the joint loss value is minimized, thus obtaining the two-branch mutual learning segmentation model.

[0011] Optionally, the formula for calculating the joint loss function is: ; in, For the joint loss value, For dynamic weighting coefficients, For the prediction results of the student model, For the prediction results of the teacher model, This is a soft loss value. This is the hard loss value; Optionally, the formula for calculating soft loss is: ; in, This is a soft loss value. n The number of training samples, For the current model in the th i Predicted probability distribution for 1 sample For the target model in the first i The predicted probability distribution for each sample.

[0012] Alternatively, the formula for calculating hard loss is: ; in, For hard losses, Cross-entropy loss is used for Dice loss, and Focal loss is used for Focal loss.

[0013] Optionally, the formula for calculating cross-entropy loss is: ; in, For cross-entropy loss, For batch size, n This represents the total number of pixels in the training samples. For the first j The first sample i The real label of each pixel For the first jThe first sample i The predicted probability of each pixel; The formula for calculating Dice loss is: ; in, For the first i The actual label value of each pixel. For the first i The predicted probability value of each pixel; The formula for calculating Focal loss is: ; in, C The total number of categories, For the first c Class weight coefficient, For the first i The pixel belongs to the first c Predicted probability of class For the first i The pixel belongs to the first c The true label of the class, For focusing parameters; categories include vascular and non-vascular.

[0014] This specification provides a blood vessel image segmentation device based on deep mutual learning, comprising: The first acquisition module is used to acquire the blood vessel image to be segmented; The second acquisition module is used to acquire the dual-branch mutual learning segmentation model. The dual-branch mutual learning segmentation model includes a parallel Attention_Unet_EfficientNetV2 model and a Base_Unet_EfficientNet model. The dual-branch mutual learning segmentation model performs feature alignment and error correction between the Attention_Unet_EfficientNetV2 model and the Base_Unet_EfficientNet model through collaborative training. The Attention_Unet_EfficientNetV2 model is built by incorporating an attention mechanism on the basis of the traditional U-Net model. The encoder of the Base_Unet_EfficientNet model is the encoder of EfficientNet-B5, and the decoder is the decoder of the traditional U-Net model. Feature transfer is achieved between the encoder and decoder through skip connections. The segmentation module is used to segment blood vessel images using a two-branch mutual learning segmentation model to obtain the blood vessel contours in the image to be segmented.

[0015] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described deep mutual learning-based blood vessel image segmentation method.

[0016] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described deep mutual learning-based blood vessel image segmentation method.

[0017] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: The deep mutual learning-based blood vessel image segmentation method provided in this specification includes a dual-branch mutual learning segmentation model comprising a parallel Attention_Unet_EfficientNetV2 model and a Base_Unet_EfficientNet model. This dual-branch structure enables the modeling of blood vessel images from different feature dimensions, receptive fields, and attention weights, achieving a complementary balance between strong expressive power and high structural stability, effectively improving the overall segmentation accuracy of blood vessels. The dual-branch mutual learning segmentation model performs feature alignment and error correction between the Attention_Unet_EfficientNetV2 model and the Base_Unet_EfficientNet model through collaborative training. The Attention_Unet_EfficientNetV2 model is constructed by integrating an attention mechanism on top of the traditional U-Net model. The Base_Unet_EfficientNet model uses an EfficientNet-B5 encoder and a traditional U-Net decoder, with feature transfer between the encoder and decoder achieved through skip connections. The dual-branch mutual learning segmentation model is used to segment the blood vessel image to be segmented, obtaining the blood vessel contours within the image.

[0018] This invention enhances fine vein extraction through attention mechanisms, preserves details through skip connections, and reduces fragmentation and missegmentation through bi-branch mutual learning. It can restore the vascular contour more completely, continuously, and accurately, especially improving the segmentation effect of small blood vessels, marginal blood vessels, and intersecting blood vessel regions, thereby improving the segmentation accuracy of vascular images. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0020] Figure 1This document provides a schematic flowchart of a blood vessel image segmentation method based on deep mutual learning. Figure 2 This is a schematic diagram of the network architecture of the dual-branch mutual learning segmentation model provided by the present invention; Figure 3 A schematic diagram of the Attention_Unet_EfficientNetV2 model provided by this invention; Figure 4 This is a schematic diagram of the Base_Unet_EfficientNet model provided by the present invention; Figure 5 The comparative experimental segmentation effect diagram provided by this invention; Figure 6 These are experimental segmentation results on different datasets provided by this invention; Figure 7 The ablation experiment segmentation effect diagram provided by this invention; Figure 8 This is a diagram showing the segmentation effect of the generalization experiment provided by the present invention; Figure 9 This is a visualization diagram of Grad-CAM provided by the present invention; Figure 10 This is a schematic diagram of a blood vessel image segmentation device based on deep mutual learning, as provided in this specification. Figure 11 This specification provides a schematic diagram of a computer device for implementing a deep mutual learning-based blood vessel image segmentation method. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0022] Devices such as desktop computers, servers, and laptops are capable of executing the solutions described in this manual. For ease of explanation, the following description will focus on servers as the primary execution method.

[0023] Currently, deep learning-based segmentation networks have become the mainstream approach, such as U-Net and its improved versions (e.g., UNet++, UNet 3+, Double U-Net). However, these methods still have the following limitations when dealing with complex vascular structures:

[0024] (1) It is not effective in segmenting small, low-contrast blood vessels, and problems such as blurred boundaries and broken branches are easily encountered.

[0025] (2) Most of them rely on training with single-modal data, and their performance degrades in cross-modal scenarios, with limited generalization ability.

[0026] (3) Existing attention mechanisms are still insufficient in multi-scale feature fusion and have limited ability to model complex vascular structures.

[0027] To address the shortcomings of existing technologies, this invention aims to propose a blood vessel image segmentation algorithm based on deep mutual learning. By constructing a dual-branch network and introducing an efficient multi-scale attention module, the following problems are solved: (1) For complex vascular structures (such as small blood vessels and low-contrast blood vessels), this invention utilizes an efficient multi-scale attention module (EMA) to model spatial and channel dependencies simultaneously through a lightweight structure, thereby achieving multi-scale information fusion and significant region enhancement. By adopting a channel grouping strategy and calculating spatial-channel joint attention weights, the feature response of key regions such as vascular boundaries is dynamically enhanced, and background noise is suppressed.

[0028] (2) To address the problem that current methods have weak generalization ability in cross-modal scenarios, this invention designs a dual-branch mutual learning network and combines it with an efficient multi-scale attention module to achieve bidirectional collaborative optimization of the feature space, thereby enhancing the feature diversity and expression stability of the model.

[0029] (3) To address the shortcomings of existing methods that focus on multi-scale feature fusion, this method introduces a hierarchical attention mechanism at the skip connections in the network. This mechanism performs spatial recalibration of the decoder features before feature concatenation, optimizing the multi-scale feature fusion effect. The model output generates pixel-level classification results through 1×1 convolution, and bilinear interpolation is used to maintain consistency between input and output sizes. This optimizes multi-scale feature fusion and enhances the model's ability to perceive vascular morphological details and topological structures.

[0030] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0031] Figure 1 This is a flowchart illustrating a blood vessel image segmentation method based on deep mutual learning as described in this specification, which specifically includes the following steps: S101: Obtain the image of the blood vessel to be segmented.

[0032] The present invention relates to vascular images to be segmented, such as coronary artery images and retinal images.

[0033] S102: Obtain the dual-branch mutual learning segmentation model; the dual-branch mutual learning segmentation model includes a parallel Attention_Unet_EfficientNetV2 model and a Base_Unet_EfficientNet model; the dual-branch mutual learning segmentation model performs feature alignment and error correction between the Attention_Unet_EfficientNetV2 model and the Base_Unet_EfficientNet model through collaborative training; the Attention_Unet_EfficientNetV2 model is built by incorporating an attention mechanism on the basis of the traditional U-Net model; the encoder of the Base_Unet_EfficientNet model is the encoder of EfficientNet-B5, and the decoder is the decoder of the traditional U-Net model. Feature transfer is achieved between the encoder and decoder through skip connections.

[0034] In an exemplary embodiment, the Attention_Unet_EfficientNetV2 model includes an encoder and a decoder; the encoder and decoder achieve feature transfer through cross-layer skip connections, and each skip connection integrates an attention gating module; the attention gating module reconstructs the spatial and channel weights of the shallow features transferred by the encoder through adaptive average pooling, grouped convolution and feature fusion.

[0035] The encoder includes multiple feature extraction layers connected in series; each feature extraction layer includes two sets of convolutional layers, batch normalization layers, and linear activation layers connected in series; each set of convolutional layers, batch normalization layers, and linear activation layers in the encoder includes a downsampling module between them.

[0036] The decoder consists of multiple feature decoding layers connected in series and a 1×1 convolutional layer; each feature decoding layer includes multiple sets of convolutional layers, batch normalization layers and linear activation layers connected in series; each of the feature decoding layers connected in series in the decoder includes an upsampling module.

[0037] The Base_Unet_EfficientNet model includes an encoder and a decoder.

[0038] The Base_Unet_EfficientNet model achieves feature transfer between the encoder and decoder through cross-layer skip connections.

[0039] The encoder of the Base_Unet_EfficientNet model consists of multiple feature extraction layers connected in series; each feature extraction layer includes two sets of convolutional layers, batch normalization layers, and linear activation layers connected in series; each set of convolutional layers, batch normalization layers, and linear activation layers in the encoder includes a downsampling module between them.

[0040] The decoder of the Base_Unet_EfficientNet model consists of multiple feature decoding layers connected in series and a 1×1 convolutional layer; each feature decoding layer includes multiple sets of convolutional layers, batch normalization layers and linear activation layers connected in series; each of the concatenated feature decoding layers in the decoder includes an upsampling module.

[0041] Specifically, Figure 2 This is a schematic diagram of the dual-branch mutual learning segmentation model provided by the present invention, as shown below. Figure 2 As shown, the dual-branch mutual learning segmentation model consists of two complementary branches: Attention_Unet_EfficientNetV2 and Base_Unet_EfficientNet. Co-training promotes mutual learning between the two models, with feature alignment and error correction performed alternately during training. Simultaneously, an interactive learning signal is constructed using a joint loss function (dml_cm_kl_loss), prompting the network to capture the feature patterns learned by the other.

[0042] S103: The blood vessel image to be segmented is segmented using a two-branch mutual learning segmentation model to obtain the blood vessel contour in the blood vessel image to be segmented.

[0043] In an exemplary embodiment, the Attention_Unet_EfficientNetV2 model further includes a classification mapping convolutional layer and a bilinear interpolation recovery module; the Base_Unet_EfficientNet model further includes a classification mapping convolutional layer and a scale alignment module; the blood vessel image to be segmented is segmented using a two-branch mutual learning segmentation model to obtain the blood vessel contour in the blood vessel image to be segmented, specifically including: inputting the blood vessel image to be segmented into the Attention_Unet_EfficientNetV2 model, performing multi-level downsampling through multiple feature extraction layers sequentially connected in the encoder to obtain the first multi-scale encoded features; performing feature fusion and decoding on the first multi-scale encoded features through a skip connection and upsampling module guided by the attention gating module to obtain the first decoded output feature Z1; inputting the first decoded output feature Z1 into the classification mapping convolutional layer for channel dimension compression processing to obtain the first branch prediction probability map. The first branch prediction probability map P1 is processed by a bilinear interpolation recovery module to obtain a first branch processing result with the same size as the blood vessel image to be segmented. The blood vessel image to be segmented is input into the Base_Unet_EfficientNet model, and the second multi-scale encoded features of the blood vessel image to be segmented are extracted through multiple feature extraction layers of the Base_Unet_EfficientNet model. The second multi-scale encoded features are restored and decoded through a skip connection in the form of direct concatenation, an upsampling module, and a multi-layer feature decoding layer to obtain the second decoded output feature Z2. The second decoded output feature Z2 is input into a classification mapping convolutional layer for class probability mapping processing to obtain the second branch prediction probability map P2. The second branch prediction probability map P2 is input into a scale alignment module for global size restoration processing to obtain the second branch processing result. The first branch processing result and the second branch processing result are fused by dual-branch inference to obtain the blood vessel contour in the blood vessel image to be segmented.

[0044] Specifically, both the first and second branch processing results are blood vessel image segmentation maps, i.e., the results of independent segmentation by the two networks. Based on the dynamic weight coefficients determined during the training phase, the first and second branch processing results are weighted and fused to obtain the blood vessel contour in the blood vessel image to be segmented.

[0045] In an exemplary embodiment, the encoder of the Attention_Unet_EfficientNetV2 model includes a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, a fourth feature extraction layer, and a fifth feature extraction layer. After multi-level downsampling through multiple feature extraction layers sequentially connected in the encoder, the first multi-scale encoded features are obtained. Specifically, this includes: inputting a 512×512 pixel blood vessel image to be segmented into the first feature extraction layer of the encoder, and passing it through two sets of convolutional layers, a batch normalization layer, and a linear activation layer sequentially connected in series to obtain a 256×256-channel... The first feature map has 24 channels. After downsampling, the first feature map is passed through a second feature extraction layer to obtain a second feature map with a dimension of 128×128 and 48 channels. After downsampling, the second feature map is passed through a third feature extraction layer to obtain a third feature map with a dimension of 64×64 and 64 channels. After downsampling, the third feature map is passed through a fourth feature extraction layer to obtain a fourth feature map with a dimension of 32×32 and 160 channels. The fourth feature map is then passed through a fifth feature extraction layer to obtain a first multi-scale encoded feature with a dimension of 16×16 and 1280 channels.

[0046] In an exemplary embodiment, the first multi-scale encoded features are fused and decoded by a skip connection and upsampling module guided by an attention gating module to obtain the first decoded output feature Z1. Specifically, this includes: upsampling the 16×16 multi-scale encoded features with 1280 channels to a dimension of 32×32, concatenating them with the attention-weighted fourth feature map in the channel dimension, and passing them through the first decoding layer in the decoder to obtain the first decoded feature with 160 channels; upsampling the first decoded feature and concatenating it with the attention-weighted third feature map in the channel dimension. The first feature map is concatenated with the second feature map in the decoder, resulting in a second decoded feature with 64 channels. The second decoded feature is then upsampled and concatenated with the attention-weighted second feature map in the channel dimension, and passed through the third decoder layer to obtain a third decoded feature with 48 channels. The third decoded feature is then upsampled and concatenated with the attention-weighted first feature map in the channel dimension, and passed through the fourth decoder layer to obtain a fourth decoded feature with 24 channels. The fourth decoded feature is then passed through a 1×1 convolutional layer to obtain the first decoded output feature Z1.

[0047] Specifically, Figure 3 This is a schematic diagram of the Attention_Unet_EfficientNetV2 model provided by the present invention, as shown below. Figure 3As shown, the Attention_Unet_EfficientNetV2 model consists of two parts: an encoder and a decoder. The encoder part uses the EfficientNet-V2-S structure for five-level downsampling feature extraction, including five feature extraction layers connected in series. It first receives a 512×512 pixel image as input, feeding the 512×512 pixel blood vessel image to be segmented into the encoder's first feature extraction layer. This first feature extraction layer includes two computational units (i.e., Double_CBR logic) containing a convolutional layer (Conv), a batch normalization layer (BN), and a linear activation layer (ReLU). The signal then passes through two sets of convolutional layers connected in series. The first feature map is obtained by passing a batch normalization layer and a linear activation layer, resulting in a first feature map with a dimension of 256×256 and 24 channels. After downsampling the first feature map, it is passed through a second feature extraction layer to obtain a second feature map with a dimension of 128×128 and 48 channels. After downsampling the second feature map, it is passed through a third feature extraction layer to obtain a third feature map with a dimension of 64×64 and 64 channels. After downsampling the third feature map, it is passed through a fourth feature extraction layer to obtain a fourth feature map with a dimension of 32×32 and 160 channels. Finally, the fourth feature map is passed through a fifth feature extraction layer to obtain a first multi-scale encoded feature with a dimension of 16×16 and 1280 channels.

[0048] The decoder stage achieves feature recovery through upsampling and skip connections. At each skip connection, an attention gating module (Efficient Multi-scale Attention, EMA) is integrated. This module reconstructs the spatial and channel weights of the shallow features transmitted by the encoder through adaptive average pooling, grouped convolution, and feature fusion techniques.

[0049] The first multi-scale encoded feature with a dimension of 16×16 and 1280 channels is upsampled to a dimension of 32×32 and then concatenated with the attention-weighted fourth feature map along the channel dimension. This concatenation is then passed through the first decoding layer in the decoder to obtain the first decoded feature with 160 channels. The first decoded feature is then upsampled and concatenated with the attention-weighted third feature map along the channel dimension. This concatenation is then passed through the second decoding layer in the decoder to obtain the second decoded feature with 64 channels. The second decoded feature is then upsampled and concatenated with the attention-weighted second feature map along the channel dimension. This concatenation is then passed through the third decoding layer in the decoder to obtain the third decoded feature with 48 channels. The third decoded feature is then upsampled and concatenated with the attention-weighted first feature map along the channel dimension. This concatenation is then passed through the fourth decoding layer in the decoder to obtain the fourth decoded feature with 24 channels. The fourth decoded feature is then passed through a 1×1 convolutional layer to obtain the first decoded output feature Z1.

[0050] Output layer: Through a 1×1 convolutional layer (Conv 1×1) and bilinear interpolation, the first branch prediction probability map P1 with a size of 512×512 and 2 channels is output. Figure 3 The medium gray line represents a skip connection path with attention weight intervention.

[0051] The Attention_Unet_EfficientNetV2 model is an improved U-Net architecture that incorporates the attention mechanism EMA. It builds an encoder-decoder structure based on EfficientNetV2 and enhances the ability to perceive blood vessel boundaries through the attention mechanism.

[0052] Figure 4 This is a schematic diagram of the Base_Unet_EfficientNet model provided by the present invention, as shown below. Figure 4 As shown, the Base_Unet_EfficientNet model adopts the same encoder-decoder architecture, using efficientNet-B5 as the backbone network and receiving a 512×512 pixel image input. Similar to the Attention_Unet_EfficientNetV2 model, the encoder stage also adopts a five-level cascaded structure, passing through the first feature extraction layer to the fifth feature extraction layer in sequence, gradually downsampling the input image, and finally generating a second multi-scale encoded feature with a dimension of 32×32 and 2048 channels in the central region. After passing through two convolutional layers, a batch normalization layer, and a linear activation layer in series, a first feature map with a dimension of 256×256 and 24 channels is obtained. This first feature map is then downsampled and passed through a second feature extraction layer to obtain a second feature map with a dimension of 128×128 and 40 channels. This second feature map is then downsampled again and passed through a third feature extraction layer to obtain a third feature map with a dimension of 64×64 and 64 channels. This third feature map is then downsampled again and passed through a fourth feature extraction layer to obtain a fourth feature map with a dimension of 32×32 and 176 channels. Finally, this fourth feature map is passed through a fifth feature extraction layer to obtain a second multi-scale encoded feature map with a dimension of 32×32 and 2048 channels. The intermediate feature maps generated by each feature extraction layer have 24, 40, 64, 176, and 2048 channels, respectively. Gray connecting lines represent skip connections.

[0053] The first multi-scale encoded feature with dimensions 32×32 and 2048 channels is upsampled to a dimension of 64×64 and then concatenated with the fourth feature map in the channel dimension. This concatenation is then passed through the first decoding layer in the decoder to obtain the first decoded feature with 176 channels. The first decoded feature is then upsampled and concatenated with the third feature map in the channel dimension. This concatenation is then passed through the second decoding layer in the decoder to obtain the second decoded feature with 64 channels. The second decoded feature is then upsampled and concatenated with the second feature map in the channel dimension. This concatenation is then passed through the third decoding layer in the decoder to obtain the third decoded feature with 40 channels. The third decoded feature is then upsampled and concatenated with the first feature map in the channel dimension. This concatenation is then passed through the fourth decoding layer in the decoder to obtain the fourth decoded feature with 24 channels. The fourth decoded feature is then passed through a 1×1 convolutional layer to obtain the first decoded output feature Z1.

[0054] The Base_Unet_EfficientNet model is an improved U-Net structure based on EfficientNet-B5. The encoder directly reuses the first few modules of EfficientNet-B5 for feature extraction, and the decoder gradually restores the spatial resolution through four upsampling operations and introduces skip connections to fuse multi-scale features.

[0055] In an exemplary embodiment, the training process of the two-branch mutual learning segmentation model specifically includes: Sample blood vessel images are acquired and their contours are annotated to obtain a training set. In the first training phase, the Attention_Unet_EfficientNetV2 and Base_Unet_EfficientNet models in the dual-branch mutual learning segmentation model are trained based on the training set until their hard losses are minimized. In the second training phase, based on the training set, a soft loss generated by knowledge distillation is introduced, and a cosine growth strategy is used to gradually increase the soft loss's proportion in the total loss. The weight coefficients are used to train the two-branch mutual learning segmentation model until it smoothly transitions from independent training to a collaborative distillation state. In the third training stage, based on the training set, the weights between the hard and soft losses are fixed, and the temperature parameter gradually decreases with each training round. The predicted probability distribution is gradually sharpened to force the two-branch mutual learning segmentation model to focus on the key anatomical features of blood vessels. In the fourth training stage, based on the training set, the parameters of the Attention_Unet_EfficientNetV2 and Base_Unet_EfficientNet models are adjusted until the joint loss value is minimized, thus obtaining the two-branch mutual learning segmentation model.

[0056] Specifically, to optimize knowledge transfer efficiency, a dynamic adjustment strategy was designed: mutual learning loss was disabled in the early training stages to stabilize basic feature extraction; cosine annealing was used in the middle stages to gradually increase the weight of the KL loss; and a fixed ratio was maintained in the later stages to balance the intensity of the supervision signal and knowledge distillation. The temperature parameter gradually decreased with each training epoch to progressively sharpen the probability distribution and focus on key features. During backpropagation, each model independently updated its parameters based on the combined loss to achieve complementary optimization of the feature space. Ultimately, this resulted in an enhanced segmentation system that both preserved individual characteristics and shared advantageous features.

[0057] The formula for calculating the joint loss function is formula (1): (1); in, For the joint loss value, For dynamic weighting coefficients, For the prediction results of the student model, This represents the prediction results from the teacher model.

[0058] The formula for calculating soft loss is formula (2): (2); in, This is a soft loss value. n The number of training samples, For the current model in the th i Predicted probability distribution for samples For the target model in the first i The predicted probability distribution for each sample.

[0059] The formula for calculating hard loss is formula (3): (3); in, Cross-entropy loss is used for Dice loss, and Focal loss is used for Focal loss.

[0060] The formula for calculating CE Loss is formula (4): (4); in, For batch size, n This represents the total number of pixels in the training samples. For the first j The first sample i The real label of each pixel For the first j The first sample i The predicted probability of each pixel.

[0061] The formula for calculating Dice Loss is formula (5): (5); in, N Total number of pixels For the first i The actual label value of each pixel. For the first i The predicted probability value for each pixel.

[0062] The formula for calculating Focal Loss is formula (6): (6); in, C The total number of categories, For the first c Class weight coefficient, For the first i The pixel belongs to the first c Predicted probability of class For the first i The pixel belongs to the first c The true label of the class, For focusing parameters.

[0063] In one exemplary embodiment, the present invention performed comparative experimental analysis, experimental analysis on different datasets, ablation experimental analysis, generalization experimental analysis, and feature map visualization analysis.

[0064] In comparative experimental analysis, this invention was tested on four publicly available vascular image datasets to verify the model's universality. The cardiovascular data primarily came from the CAGData dataset constructed by a hospital and the publicly available DB_Angiograms_134 dataset. The CAGData dataset contains 184 high-resolution coronary angiography images (992×992 pixels, PNG format), all with manually annotated vascular contours, providing a reliable basis for coronary artery stenosis detection and plaque analysis. The DB_Angiograms_134 dataset includes 134 coronary artery images (300×300 pixels, PGM format), each with a real vascular mask annotated by a cardiologist, suitable for model training and validation. Combining these two datasets can support precision diagnosis and treatment research of cardiovascular diseases and provide diverse clinical scenario samples for AI algorithm development.

[0065] The fundus and retinal data primarily utilize the publicly available DRIVE and CHASE_DB1 datasets. The DRIVE dataset contains 40 color fundus images (565×584 pixels, TIF format) and expert-segmented vessel annotations (GIF format). Its standardized annotations and clear structural features make it a benchmark set for retinal vessel segmentation algorithms. The CHASE_DB1 dataset contains 28 high-resolution fundus images (999×960 pixels, JPG format), each accompanied by ophthalmologist-annotated vessel masks (PNG format), making it particularly suitable for early screening studies of diabetic retinopathy. The open sharing of these datasets not only promotes the development of fundus image analysis technology but also provides crucial support for cross-institutional collaboration and the evaluation of algorithm generalization capabilities.

[0066] This invention performs data augmentation on the four datasets mentioned above, and the information of the four datasets after data preprocessing is shown in Table 1.

[0067] Table 1 shows the four datasets after data preprocessing. This invention's model is implemented based on the PyTorch framework. All experiments were conducted in a server environment equipped with an Intel Xeon Platinum 8352V CPU and an NVIDIA GPU with 32 GB of video memory. Training was performed using the Adam optimizer, with an initial learning rate of 0.0001, a weight decay factor of 0.0005, and a batch size of 4. This configuration achieved a good balance between memory usage and training stability.

[0068] To comprehensively evaluate the segmentation performance of the model, this invention employs six commonly used metrics: Accuracy (Acc), Intersection over Union (IoU), Dice Coefficient, Sensitivity (SE), Specificity (SP), and Area Under the ROC Curve (AUC). Accuracy measures the proportion of correctly classified pixels, reflecting the model's overall recognition ability. IoU and Dice are used to assess the overlap between the predicted region and the ground truth annotation; the former is sensitive to boundary differences, while the latter is more suitable for evaluating the overlap accuracy of small targets. Sensitivity represents the proportion of correctly identified blood vessel pixels out of the actual blood vessel pixels, measuring the sufficiency of blood vessel detection. Specificity represents the proportion of correctly identified background pixels, evaluating the model's ability to suppress false positives. AUC, by calculating the relationship between the true positive rate and the false positive rate at different thresholds, reflects the model's robustness and discriminative ability in global classification tasks.

[0069] This invention selects several mainstream segmentation models, such as UNet3+ and ResUNet, as benchmarks and systematically compares them with two proposed models after incorporating Deep Mutual Learning (DML) (Attention_Unet_EfficientNetV2 denoted as Net1, and Base_Unet_EfficientNet denoted as Net2), as shown in Table 2. On the CAGData dataset, mainstream models such as UNet3+ and ResUNet exhibit high accuracy (Acc≥0.973) and Dice coefficients (0.718–0.903), but their -10-IoU (0.623–0.841) and specificity (SP≤0.798) show significant fluctuations, indicating limited generalization ability for complex boundaries. In contrast, on the DRIVE dataset, Net1+DML and Net2+DML, which incorporate deep mutual learning, significantly outperform the benchmark models. For example, Net1+DML's IoU (0.862) and Dice (0.921) are improved by about 8.9% and 5.4% respectively compared to UNet3+ (IoU=0.791, Dice=0.874). At the same time, its sensitivity (SE=0.843) and AUC (0.917) are also better, indicating that DML effectively improves the model's ability to recognize fine structures by enhancing feature interaction.

[0070] Table 2 Comparative Experiments Further analysis of the overall performance on the DRIVE dataset reveals that while Net2+DML's AUC (0.897) and IoU (0.816) are slightly lower than those in the literature (AUC=0.978), its specificity (SP=0.801) and sensitivity (SE=0.820) are more balanced, indicating a better trade-off between reducing false positives and false negatives. Furthermore, the Acc values ​​of both Net1+DML and Net2+DML remain stable at 0.984 and 0.955, respectively, significantly higher than most benchmark models on the same dataset (e.g., ResUNet Acc=0.915). This result validates that the DML strategy optimizes network parameters through bidirectional knowledge distillation, enabling the model to maintain high accuracy while enhancing the consistency of feature representation.

[0071] Figure 5 A comparison diagram of the segmentation results provided by this invention, such as Figure 5 As shown, Net1+DML and Net2+DML achieve multi-dimensional performance improvements through a deep mutual learning mechanism. On the one hand, DML enhances the model's ability to capture detailed features, which is reflected in the significant optimization of IoU and Dice metrics. On the other hand, it balances sensitivity and specificity through collaborative training, preventing a single model from becoming biased towards a particular metric. Compared with models such as UNet3+ and ResUNet, the method proposed in this invention achieves superior overall performance while maintaining high accuracy, making it valuable for applications in medical image segmentation tasks where edge accuracy and generalization ability must be balanced. Experimental results demonstrate that the DML strategy effectively improves model performance, providing a new direction for future optimization research.

[0072] As shown in Table 3, the experimental results on different datasets reveal that, on four medical imaging datasets (CAGData and DB_Angiograms_134 for cardiovascular data; DRIVE and CHASE_DB1 for retinal blood vessels), Attention_Unet_EfficientNetV2 (Net1) and Base_Unet_EfficientNet (Net2) exhibit differentiated performance improvements after incorporating the Deep Mutual Learning (DML) strategy. On the CAGData dataset, Net1+DML increased Acc from 0.983 to 0.984, IoU and Dice by 0.005 and 0.003 respectively, while AUC increased from 0.904 to 0.917, indicating that DML effectively enhanced the model's ability to capture complex features.

[0073] Table 3 Experiments on different datasets Net2+DML also improved Acc and IoU by 0.002 and 0.008 respectively, but SP decreased slightly (0.813→0.847), suggesting that DML may have sacrificed some specificity through feature sharing. In the DRIVE dataset, Net2+DML performed particularly well, with Acc improving from 0.951 to 0.955, and IoU and Dice improving by 0.012 and 0.008 respectively, but SE and AUC decreased slightly, reflecting the trade-off between sensitivity and robustness. In the CHASE_DB1 dataset, although Net2+DML showed a slight improvement in Acc and IoU, AUC decreased from 0.898 to 0.893, indicating that DML had limited impact on the overall generalization ability of the model. Segmentation performance comparison: Figure 6 As shown in the experimental results, Net1 significantly improves AUC and IoU on the CAGData and DRIVE datasets with DML; for Acc and Dice, Net2+DML achieves higher overall performance on DRIVE. However, introducing DML may slightly decrease SP or AUC, requiring a trade-off based on task requirements. Overall, Net1+DML outperforms in AUC on most datasets, while Net2+DML shows greater potential in complex scenarios like DRIVE.

[0074] In the ablation experiment analysis, Table 4 shows the ablation results of UNet_EfficientNet (Model 1) and UNet_EfficientNetV2 (Model 2). The experiment used ce_dice_focal (denoted as A) as the loss and EMA (denoted as B) as the attention module. In the initial stage, both models were trained using only CE loss and standard attention as a control. The results show that ce_dice_focal is slightly better in terms of Dice and IoU, indicating that the composite loss alleviates class imbalance and improves small vessel recognition. The accuracy and specificity are further improved after introducing EMA, indicating that it stabilizes the gradient and enhances feature robustness. The performance is optimal when both are used simultaneously, verifying the complementary effect. In summary, both the loss design and the attention mechanism significantly improve model performance. The ablation experiment segmentation results are as follows: Figure 7 ,like Figure 7 As shown, ablation experiments revealed that the ce_dice_focal loss function and the EMA attention mechanism both have a positive effect on improving model performance.

[0075] Table 4 Ablation Experiment In the generalization experiment analysis, Table 5 shows the experimental results of introducing the attention mechanism EMA on four base models: UNet, Base_UNet_ResNet101, Base_UNet_EfficientNet, and Base_UNet_EfficientNetV2. The segmentation results are as follows: Figure 8 As shown, the combination of EMA and the base UNet performs particularly well on the CAGData dataset: Acc increases from 0.984 to 0.985, segmentation metrics IoU and Dice increase by 0.008 and 0.005 respectively, while SP significantly improves from 0.848 to 0.869. This result fully demonstrates that EMA, through its dynamic feature weighting mechanism, can effectively enhance the model's focus on key regions, thereby improving segmentation accuracy while strengthening robustness to noise. Furthermore, the IoU and Dice of EMA+UNet_ResNet101 also reach 0.864 and 0.922 respectively, showcasing EMA's adaptability across different backbone networks. Notably, after introducing EMA, the AUC of Base_UNet_EfficientNetV2 further climbs from 0.910 to 0.914, indicating that EMA helps the model more accurately balance sensitivity and specificity, improving overall discriminative ability.

[0076] Table 5 Generalization experiments on EMA In the DRIVE dataset, the gains brought by EMA are also significant. When EMA is combined with UNet, Acc increases from 0.955 to 0.957, Dice improves from 0.897 to 0.903, and AUC remains consistently at a high level of 0.913. These data demonstrate that EMA effectively improves the consistency of the model in segmenting complex vascular structures through the dynamic integration of global features. In addition, after adding EMA to UNet_EfficientNetV2, the sensitivity (SE) increases significantly from 0.863 to 0.876, further confirming the potential of EMA in improving the ability to capture local details. Although the SP index of some models fluctuates, the stable improvement of EMA in core metrics (such as Acc, Dice, and AUC) still highlights its value as a general optimization strategy.

[0077] Overall, the introduction of the EMA mechanism can improve the performance of models in medical image segmentation tasks. It not only optimizes segmentation accuracy (e.g., IoU, Dice) through dynamic feature weighting, but also enhances the model's robustness to noise (e.g., SP) and overall discriminative ability (e.g., AUC). These results demonstrate that EMA, as a flexible feature enhancement strategy, can effectively adapt to models with different architectures, providing reliable technical support for medical image analysis tasks. Future research can further explore the synergistic design of EMA with other advanced modules to unlock its application value in more scenarios.

[0078] In feature map visualization analysis, such as Figure 9 As shown, it is clear that the EMA module focuses specifically on detailed areas closely related to the task. This design demonstrates the model's superior ability to capture key information and further enhances its analytical capabilities for complex scenarios. This visualization provides a more intuitive understanding of the EMA module's working principle and offers a deeper insight into its crucial role in the model.

[0079] When applying the deep mutual learning-based blood vessel image segmentation method provided in this manual, it is not necessary to... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this manual does not impose any restrictions on it.

[0080] The above describes one or more embodiments of a deep mutual learning-based blood vessel image segmentation method provided in this specification. Based on the same approach, this specification also provides a corresponding deep mutual learning-based blood vessel image segmentation device, such as... Figure 10 As shown.

[0081] Figure 10 A schematic diagram of a blood vessel image segmentation device based on deep mutual learning provided in this specification includes: The first acquisition module 1001 is used to acquire the blood vessel image to be segmented.

[0082] The second acquisition module 1002 is used to acquire a two-branch mutual learning segmentation model. The two-branch mutual learning segmentation model includes a parallel Attention_Unet_EfficientNetV2 model and a Base_Unet_EfficientNet model. The two-branch mutual learning segmentation model performs feature alignment and error correction between the Attention_Unet_EfficientNetV2 model and the Base_Unet_EfficientNet model through collaborative training. The Attention_Unet_EfficientNetV2 model is built by incorporating an attention mechanism on the basis of the traditional U-Net model. The encoder of the Base_Unet_EfficientNet model is the encoder of EfficientNet-B5, and the decoder is the decoder of the traditional U-Net model. Feature transfer is achieved between the encoder and the decoder through skip connections.

[0083] The segmentation module 1003 is used to segment the blood vessel image to be segmented through a two-branch mutual learning segmentation model to obtain the blood vessel contour in the blood vessel image to be segmented.

[0084] Specific limitations regarding the deep mutual learning-based vascular image segmentation device can be found in the limitations of the deep mutual learning-based vascular image segmentation method described above, and will not be repeated here. Each module in the aforementioned deep mutual learning-based vascular image segmentation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0085] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The provided method is a blood vessel image segmentation method based on deep mutual learning.

[0086] This instruction manual also provides Figure 11 The schematic diagram of the computer device shown is as follows: Figure 11 At the hardware level, the computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above-mentioned functions. Figure 1 The provided method is a blood vessel image segmentation method based on deep mutual learning.

[0087] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0088] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A blood vessel image segmentation method based on deep mutual learning, characterized in that, include: Acquire the image of the blood vessel to be segmented; A dual-branch mutual learning segmentation model is obtained; the dual-branch mutual learning segmentation model includes a parallel Attention_Unet_EfficientNetV2 model and a Base_Unet_EfficientNet model; the dual-branch mutual learning segmentation model performs feature alignment and error correction between the Attention_Unet_EfficientNetV2 model and the Base_Unet_EfficientNet model through collaborative training; the Attention_Unet_EfficientNetV2 model is built by fusing an attention mechanism on the basis of the traditional U-Net model; the encoder of the Base_Unet_EfficientNet model is the encoder of EfficientNet-B5, and the decoder is the decoder of the traditional U-Net model, with feature transfer between the encoder and decoder achieved through skip connections; The blood vessel image to be segmented is segmented using the bi-branch mutual learning segmentation model to obtain the blood vessel contour in the blood vessel image to be segmented.

2. The blood vessel image segmentation method based on deep mutual learning as described in claim 1, characterized in that, The Attention_Unet_EfficientNetV2 model includes an encoder and a decoder; the encoder and decoder achieve feature transfer through cross-layer skip connections, and each skip connection integrates an attention gating module; the attention gating module reconstructs the spatial and channel weights of the shallow features transferred by the encoder through adaptive average pooling, grouped convolution and feature fusion. The encoder of the Attention_Unet_EfficientNetV2 model consists of multiple feature extraction layers connected in series; each feature extraction layer includes two sets of convolutional layers, batch normalization layers, and linear activation layers connected in series; each set of convolutional layers, batch normalization layers, and linear activation layers in the encoder includes a downsampling module between them. The decoder of the Attention_Unet_EfficientNetV2 model consists of multiple feature decoding layers connected in series and a 1×1 convolutional layer; Each feature decoding layer includes multiple sets of convolutional layers, batch normalization layers, and linear activation layers connected in series; each of the concatenated feature decoding layers in the decoder includes an upsampling module. The Base_Unet_EfficientNet model includes an encoder and a decoder; The Base_Unet_EfficientNet model achieves feature transfer between the encoder and decoder through cross-layer skip connections; The encoder of the Base_Unet_EfficientNet model consists of multiple feature extraction layers connected in series; each feature extraction layer includes two sets of convolutional layers, batch normalization layers, and linear activation layers connected in series; each set of convolutional layers, batch normalization layers, and linear activation layers in the encoder includes a downsampling module between them; The decoder of the Base_Unet_EfficientNet model consists of multiple feature decoding layers connected in series and a 1×1 convolutional layer; Each feature decoding layer includes multiple sets of convolutional layers, batch normalization layers, and linear activation layers connected in series; each of the concatenated feature decoding layers in the decoder includes an upsampling module.

3. The blood vessel image segmentation method based on deep mutual learning as described in claim 2, characterized in that, The Attention_Unet_EfficientNetV2 model further includes a classification mapping convolutional layer and a bilinear interpolation recovery module; the Base_Unet_EfficientNet model further includes a classification mapping convolutional layer and a scale alignment module; the process of segmenting the blood vessel image to be segmented using the dual-branch mutual learning segmentation model to obtain the blood vessel contour in the blood vessel image to be segmented specifically includes: The blood vessel image to be segmented is input into the Attention_Unet_EfficientNetV2 model, and after multiple feature extraction layers connected in series in the encoder, multi-level downsampling is performed to obtain the first multi-scale encoded features; The first multi-scale encoded feature is fused and decoded by the skip connection and upsampling module guided by the attention gating module to obtain the first decoded output feature Z1. The first decoded output feature Z1 is input into the classification mapping convolutional layer for channel dimension compression processing to obtain the first branch prediction probability map P1. The first branch prediction probability map P1 is then processed by the bilinear interpolation recovery module to obtain the first branch processing result with the same size as the blood vessel image to be segmented. The blood vessel image to be segmented is input into the Base_Unet_EfficientNet model, and the second multi-scale encoded features of the blood vessel image to be segmented are extracted through multiple feature extraction layers of the Base_Unet_EfficientNet model. By using a skip connection in the form of direct splicing, an upsampling module, and a multi-layer feature decoding layer, the second multi-scale encoded feature is restored and decoded to obtain the second decoded output feature Z2. The second decoded output feature Z2 is input into the classification mapping convolutional layer for class probability mapping processing to obtain the second branch prediction probability map P2. The second branch prediction probability map P2 is input into the scale alignment module for global size restoration processing to obtain the second branch processing result; The results of the first branch processing and the second branch processing are fused by dual-branch inference to obtain the blood vessel contour in the blood vessel image to be segmented.

4. The blood vessel image segmentation method based on deep mutual learning as described in claim 3, characterized in that, The encoder of the Attention_Unet_EfficientNetV2 model includes a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, a fourth feature extraction layer, and a fifth feature extraction layer. The first multi-scale encoded features are obtained by multi-level downsampling through multiple feature extraction layers sequentially connected in the encoder, specifically including: The 512×512 pixel blood vessel image to be segmented is input into the first feature extraction layer of the encoder. After passing through two sets of convolutional layers, batch normalization layers and linear activation layers in sequence, a first feature map with a dimension of 256×256 and a channel number of 24 is obtained. After downsampling the first feature map, it is passed through the second feature extraction layer to obtain a second feature map with a dimension of 128×128 and a number of channels of 48. After downsampling the second feature map, it is passed through the third feature extraction layer to obtain a third feature map with a dimension of 64×64 and a number of channels of 64. After downsampling, the third feature map is passed through the fourth feature extraction layer to obtain a fourth feature map with a dimension of 32×32 and a number of channels of 160. The fourth feature map is passed through the fifth feature extraction layer to obtain the first multi-scale coded feature with a dimension of 16×16 and a channel number of 1280.

5. The blood vessel image segmentation method based on deep mutual learning as described in claim 4, characterized in that, The skip connection and upsampling module guided by the attention gating module performs feature fusion and decoding on the first multi-scale encoded features to obtain the first decoded output feature Z1, specifically including: The multi-scale encoded features with dimensions of 16×16 and 1280 channels are upsampled to a dimension of 32×32 and then concatenated with the attention-weighted fourth feature map in terms of channel dimension. After passing through the first decoding layer in the decoder, the first decoded feature with 160 channels is obtained. The first decoded feature is upsampled and then concatenated with the attention-weighted third feature map in the channel dimension, and then passed through the second decoding layer in the decoder to obtain the second decoded feature with 64 channels. The second decoded feature is upsampled and then concatenated with the attention-weighted second feature map in the channel dimension. After passing through the third decoding layer in the decoder, a third decoded feature with 48 channels is obtained. The third decoding feature is upsampled and then concatenated with the attention-weighted first feature map in the channel dimension, and then passed through the fourth decoding layer in the decoder to obtain the fourth decoding feature with 24 channels. The fourth decoding feature is passed through a 1×1 convolutional layer to obtain the first decoding output feature Z1.

6. The blood vessel image segmentation method based on deep mutual learning as described in claim 1, characterized in that, The training process of the dual-branch mutual learning segmentation model specifically includes: Acquire sample blood vessel images and annotate the blood vessel contours in the sample blood vessel images to obtain a training set; In the first training phase, the Attention_Unet_EfficientNetV2 model and the Base_Unet_EfficientNet model in the dual-branch mutual learning segmentation model are trained based on the training set until the hard loss values ​​of the Attention_Unet_EfficientNetV2 model and the Base_Unet_EfficientNet model reach their minimum. In the second training phase, based on the training set, soft loss generated by knowledge distillation is introduced, and a cosine growth strategy is used to gradually increase the weight coefficient of the soft loss in the total loss. The two-branch mutual learning segmentation model is trained until the two-branch mutual learning segmentation model smoothly transitions from independent training to the cooperative distillation state. In the third training phase, based on the training set, the weights between hard loss and soft loss are fixed and the temperature parameter is gradually reduced as the training rounds increase. The prediction probability distribution is gradually sharpened to force the bi-branch mutual learning segmentation model to focus on the key anatomical features of blood vessels. In the fourth training phase, based on the training set, the parameters of the Attention_Unet_EfficientNetV2 model and the Base_Unet_EfficientNet model are adjusted until the joint loss value is minimized, thus obtaining the two-branch mutual learning segmentation model.

7. The blood vessel image segmentation method based on deep mutual learning as described in claim 6, characterized in that, The formula for calculating the joint loss function is: ； in, For the joint loss value, For dynamic weighting coefficients, For the prediction results of the student model, For the prediction results of the teacher model, This is a soft loss value. This is the hard loss value.

8. The blood vessel image segmentation method based on deep mutual learning as described in claim 6, characterized in that, The formula for calculating the soft loss is: ； in, This is a soft loss value. n The number of training samples, For the current model in the th i Predicted probability distribution for a sample number For the target model in the first i The predicted probability distribution for each sample.

9. The blood vessel image segmentation method based on deep mutual learning as described in claim 6, characterized in that, The formula for calculating the hard loss is: ； in, For hard losses, Cross-entropy loss is used for Dice loss, and Focal loss is used for Focal loss.

10. The blood vessel image segmentation method based on deep mutual learning as described in claim 9, characterized in that, The formula for calculating the cross-entropy loss is as follows: ； in, For cross-entropy loss, For batch size, n This represents the total number of pixels in the training samples. For the first j The first sample i The real label of each pixel For the first j The first sample i The predicted probability of each pixel; The formula for calculating the Dice loss is as follows: ； in, For the first i The actual label value of each pixel. For the first i The predicted probability value of each pixel; The formula for calculating the Focal loss is as follows: ； in, C The total number of categories, For the first c Class weight coefficient, For the first i The pixel belongs to the first c Predicted probability of class For the first i The pixel belongs to the first c The true label of the class, For focusing parameters; categories include vascular and non-vascular.

Images