Incremental learning-based retinal vessel segmentation and lesion detection method

Abstract

The invention discloses a retinal vessel segmentation and lesion detection method based on incremental learning. The method comprises the specific steps that firstly, an original image of a data set STARE is acquired, and preprocessing such as segmentation and data enhancement is carried out on an original retina image; then, a VGAT-Net-IL network model is constructed, the network takes a coding-decoding symmetric structure as a trunk, a visual cortex mechanism is simulated through an adaptive receptive field module to dynamically adjust a receptive field, and local details and global features of the retinal vessels are cooperatively extracted; meanwhile, a dynamic bimodal attention module is innovatively integrated, variable convolution is introduced into the dynamic bimodal attention module to adaptively adjust a sampling position, a blood vessel region is precisely focused in combination with a space and channel attention mechanism, and after the dynamic bimodal attention module, a Bayesian semantic association module is introduced to generate features containing semantic association; in order to solve the problem that old knowledge is easy to forget when a model learns new lesion features, an incremental learning technology training model is introduced. According to the method, the retinal vessel segmentation precision and the lesion detection capability are improved, and a reliable image analysis basis is provided for retinal disease diagnosis.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing technology, and specifically relates to a retinal blood vessel segmentation and lesion detection method based on incremental learning Background Art

[0002] Retinal vessel segmentation and lesion detection play a crucial role in the diagnosis of ophthalmic diseases. For diseases like diabetic retinopathy and glaucoma, early detection and assessment rely not only on accurate retinal vessel segmentation results but also on lesion detection. By analyzing the morphology, distribution, and structural changes of retinal vessels, as well as detecting the type, location, and severity of lesions, doctors can be provided with comprehensive and critical diagnostic information, assisting in the development of precise and effective treatment plans.

[0003] Currently, retinal vessel segmentation and lesion detection face numerous challenges. Traditional segmentation methods, such as those based on thresholding, edge detection, and region growing, perform poorly when processing complex retinal images. These methods struggle to cope with issues such as low contrast between vessels and background, variable vessel morphology, and image noise. They are often based on simple image features and lack a deep understanding of the complex structure and semantics of retinal vessels. This results in limited segmentation accuracy and is prone to mis-segmentation and missed segmentations.

[0004] With the development of deep learning technology, convolutional neural network (CNN)-based methods have made some progress in the field of retinal vessel segmentation. However, existing deep learning models still have some limitations. Some models have a large number of parameters and high computational complexity, resulting in long training and inference times, making them difficult to quickly deploy in actual clinical applications. Moreover, many models lack the ability to capture the multi-scale characteristics and local details of retinal vessels during feature extraction, making it impossible to fully utilize the effective information in the image, which affects segmentation accuracy.

[0005] When it comes to lesion detection, both traditional methods and some existing deep learning models suffer from shortcomings. Traditional methods often struggle to accurately identify subtle lesion features, limiting their ability to detect newly emerging or rare lesion types. Existing deep learning models are prone to "catastrophic forgetting" when faced with a constant influx of lesion data—learning new lesion features while forgetting previous knowledge. Furthermore, these models lack an understanding and utilization of semantic connections between different lesions, making it difficult to comprehensively assess complex lesions. This limits the accuracy and comprehensiveness of detection.

[0006] In order to overcome the shortcomings of the above-mentioned existing technologies, the present invention proposes a retinal vessel segmentation and lesion detection method based on incremental learning, aiming to improve the accuracy and efficiency of retinal vessel segmentation and lesion detection, and provide a more reliable image analysis basis for the accurate diagnosis of ophthalmic diseases. Summary of the Invention

[0007] The present invention provides a retinal blood vessel segmentation and lesion detection method based on incremental learning.

[0008] The purpose of the present invention is to solve many problems existing in existing retinal blood vessel segmentation and lesion detection technologies. In terms of blood vessel segmentation, it is committed to overcoming the difficulties of insufficient accuracy, weak ability to capture complex features, and low computational efficiency; in terms of lesion detection, it focuses on solving the problems that traditional methods have difficulty in accurately identifying subtle features of lesions, have limited ability to detect new lesion types, and that existing deep learning models are prone to "catastrophic forgetting" and have difficulty understanding the semantic associations between lesions. By constructing an advanced network model, the present invention can not only achieve high-precision retinal blood vessel segmentation, but also improve the accuracy and comprehensiveness of lesion detection, providing a reliable image analysis basis for accurate diagnosis and condition assessment of ophthalmic diseases.

[0009] The technical solution adopted by the present invention comprises the following steps:

[0010] (1) The present invention first implements a preprocessing strategy on the original retinal image. Specifically, a region partitioning algorithm is used to segment the original retinal image into multiple sub-regions with specific spatial scales. Subsequently, a variety of data enhancement techniques are applied to each sub-region image, including but not limited to rotation and flipping operations based on geometric transformations, to simulate retinal imaging conditions under different viewing angles.

[0011] (2) Construct the VGAT-Net-IL network model, which adopts the encoding-decoding symmetric structure as the backbone architecture;

[0012] (2a) First, the preprocessed data x in step (1) is p Input to the downsampling module to reduce the spatial resolution of the data and expand the receptive field, laying the foundation for deep feature extraction;

[0013]

[0014] Among them, x p is the preprocessed data, k1, k2 are the pooling kernel size and convolution kernel size respectively, s1, s2 are the step size of pooling and convolution operation respectively;

[0015] (2b) The downsampled feature map x d The image is then input into the adaptive receptive field module (VGA-Module). The module simulates the biological visual cortex mechanism and dynamically adjusts the size and range of the receptive field based on the input image features, thereby collaboratively extracting local details of retinal blood vessels (such as edge texture and diameter changes) and global features (such as the overall direction and distribution of blood vessels).

[0016] Its local features:

[0017] x m =ResNet(x d )=CBblock n (...(CBblock2(CBblock1(x d )))...)

[0018] Its global features, the multi-channel feature map x extracted in step (2a) d Perform spatial dimension division to generate non-overlapping image blocks of fixed size, and map them into feature vector sequences through linear projection. This encoding method embeds spatial position information into the feature dimension and preserves the relative position relationship between feature blocks.

[0019] Its global characteristics:

[0020] x p =[(x pj ·W)]+x ep

[0021] Among them, x pj is the projected patch feature matrix, W is the weight matrix, x ep is the position encoding matrix;

[0022] (x a ) i+1 =MHA-VGAP(L((x p ) i )+(x p ) i ,i=0,...N

[0023] MHA-VGAP=MultiHead(VGAP(q),VGAP(k),VGAP(v))

[0024] VGAP(q)=q w:ap ·q ap +q w:mp ·q mp

[0025] Among them, (x p ) i is the feature vector with position encoding of the input of the i-th layer, L is the layer normalization, and MHA-VGAP is the multi-head attention calculation with visually guided adaptive pooling

[0026] (x p ) i+1 =MLP(L((x a ) i+1 ))+(x a )i+1 ,i=0,1...N

[0027] The fusion of local features and global features

[0028] x m,n =w l ·x m +w g ·x n

[0029] Among them, w l ,w g The weights for adaptive dynamic allocation of local and global features respectively;

[0030] (2c) Innovatively introduce a dynamic bimodal attention module to achieve accurate representation of retinal blood vessels through a multi-dimensional feature enhancement strategy: Dynamic perception mechanism in spatial domain: A deformable convolution system based on residual learning, which dynamically adjusts the standard convolution grid p by designing an offset generation network o The sampling position is optimized through reinforcement learning, so that the model can adaptively capture the complex geometric topology of the vascular tree. It is combined with the feature non-local association enhancement unit: a spatial attention module based on non-local operations is constructed. By calculating the association weight between any two points in the feature map, the generated attention map can adaptively enhance the response of the vascular area. And channel dimension screening: a channel attention mechanism based on the squeeze-and-excite module is designed. Channel statistics are generated through global average pooling (GAP), and channel weights are generated through a two-layer fully connected network.

[0031] Its variable convolution calculation process:

[0032]

[0033] Among them, W k is the convolution kernel weight, X m,n is the input feature, p0 is the center position, p k is the regular network position, Δ pk is the offset for learning;

[0034] Its edge feature calculation process:

[0035]

[0036] Among them, σ is the Sigmoid function, G x , G y They are gradient operations in vertical and horizontal directions respectively;

[0037] The spatial attention weight calculation process:

[0038] W s=Sigmoid(Conv(GAP([X l ;X e ])))

[0039] Its channel attention weight calculation process:

[0040] W c =Sigmoid(FC(GAP(X l )))

[0041] Its adaptive feature fusion calculation process:

[0042]

[0043] (3) Introducing incremental learning technology to train the model. In the first stage, VGAT-Net-IL is trained using fundus image datasets such as STARE, so that the model can learn the feature representation of retinal blood vessel segmentation and basic lesions (such as microaneurysms), and output accurate blood vessel segmentation masks and lesion classification results. In the second stage, when new lesion data (such as exudate-related data) are introduced, continuous learning loss is used to optimize the model. The continuous learning loss consists of blood vessel segmentation loss, old lesion classification loss, new lesion classification loss, and mutual distillation loss;

[0044] Vessel segmentation loss (L seg ): It is used to measure the difference between the retinal vessel segmentation results predicted by the model and the actual vessel area.

[0045]

[0046] Where N is the total number of pixels, y i is the true label of pixel i, is the probability that the model predicts that pixel i belongs to a blood vessel.

[0047] Old lesion classification loss (L class-old ): Maintain the model's recognition accuracy for old lesions that have been learned.

[0048]

[0049] Where M is the number of old lesion samples, C old is the number of old lesion categories, y jk Indicates whether sample j belongs to old lesion category k, p jk is the probability that the model predicts that sample j belongs to old lesion category k.

[0050] New lesion classification loss (L class-new ): Guide the model to learn the newly introduced lesion type features.

[0051]

[0052] Where N is the number of new lesion samples, C new is the number of new lesion categories, y lm Indicates whether the sample l belongs to the new lesion category m, p lm is the probability that the model predicts that sample l belongs to the new lesion category m.

[0053] Distillation loss (L md ): Calculate the joint probability distribution of new and old lesions based on the Bayesian rule, aiming to learn the structural and semantic dependencies between new and old lesions and establish associations between different lesions.

[0054]

[0055] Among them, C old For old lesion category, C new is the lesion category, p(c old ,c new ) is the prior joint probability of the new and old lesion categories, p(z|c old ,c new ) is the likelihood function of feature z given the new and old lesion categories, p(c old ) is the prior probability of the old lesion class, and p(z) is the probability of feature z.

[0056] (4) The trained VGAT-Net-IL network model is applied to the test image to segment the retinal blood vessels and detect lesions, and finally obtain high-precision retinal blood vessel segmentation results and lesion detection results.

[0057] The beneficial effects of the present invention are as follows: the adaptive receptive field module simulates the biological visual cortex mechanism, dynamically adjusts the receptive field, and collaboratively extracts local details of retinal blood vessels (such as edge texture, diameter changes) and global features (such as overall direction, distribution status); the adaptive dual local attention module uses variable convolution, spatial and channel attention mechanisms to accurately focus on the blood vessel area, comprehensively and accurately capture blood vessel features, improve segmentation accuracy, reduce missegmentation and missed segmentation, introduce incremental learning technology, and combine it with a staged training model method. It uses a continuous learning loss composed of blood vessel segmentation loss, old lesion classification loss, new lesion classification loss and mutual distillation loss to optimize the model. It not only maintains the model's recognition accuracy for learned old lesions, but also guides the model to accurately learn the newly introduced lesion type features, and establishes structural and semantic dependency relationships between new and old lesions, thereby improving the model's understanding and processing capabilities for complex lesion conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart for implementing the present invention.

[0059] Figure 2 This is a network model diagram of the retinal blood vessel segmentation and lesion detection method based on incremental learning in the present invention.

[0060] Figure 3 Schematic diagram of retinal blood vessel segmentation according to the present invention.

[0061] Figure 4 Schematic diagram of the adaptive receptive field module fusion algorithm (VGAFF) module of the present invention.

[0062] Figure 5 Schematic diagram of the adaptive dual local attention module (ADLA) of the present invention. DETAILED DESCRIPTION

[0063] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0064] The process of the present invention is as follows Figure 1 The specific implementation steps are as follows:

[0065] (1) Download the fundus image dataset STARE from the official website. The dataset contains 20 retinal images and 10 images containing fundus lesions with a resolution of 700 × 605 pixels. It is divided into a training set (15 images) and a validation set (5 images).

[0066] (2) The present invention first implements a preprocessing strategy on the original retinal image. Specifically, a region partitioning algorithm is used to segment the original retinal image into multiple sub-regions with specific spatial scales. Subsequently, a variety of data enhancement techniques are applied to each sub-region image, including but not limited to rotation and flipping operations based on geometric transformations, to simulate retinal imaging conditions under different viewing angles.

[0067] (3) Construct the VGAT-Net-IL network model, which adopts the encoding-decoding symmetric structure as the backbone architecture;

[0068] (3a) First, the preprocessed data x in step (1) is p The input is fed into the downsampling module, which uses three consecutive convolutional layers with a stride of 2, a kernel size of 3×3, and channels of 32, 64, and 128, downsampling the input image size from 700×605 to 88×76. This reduces the spatial resolution of the data while expanding the receptive field, allowing subsequent modules to capture more macroscopic image information.

[0069]

[0070] Among them, x p is the preprocessed data, k1, k2 are the pooling kernel size and convolution kernel size respectively, s1, s2 are the step size of pooling and convolution operation respectively;

[0071] (3b) The downsampled feature map x d The image is then input into the adaptive receptive field module (VGA-Module). The module simulates the biological visual cortex mechanism and dynamically adjusts the size and range of the receptive field based on the input image features, thereby collaboratively extracting local details of retinal blood vessels (such as edge texture and diameter changes) and global features (such as the overall direction and distribution of blood vessels).

[0072] Its local features use three residual blocks, each of which contains two 3×3 convolutional layers with channel numbers of 128, 256, and 512 respectively:

[0073] x m =ResNet(x d )=CBblock n (...(CBblock2(CBblock1(x d )))...)

[0074] Its global features, the multi-channel feature map x extracted in step (2a) d Perform spatial dimension division to generate non-overlapping image blocks of fixed size, and map them into feature vector sequences through linear projection. This encoding method embeds spatial position information into the feature dimension and preserves the relative position relationship between feature blocks.

[0075] x p =[(x pj ·W)]+x ep

[0076] Among them, x pj is the projected patch feature matrix, W is the weight matrix, x ep is the position encoding matrix;

[0077] (x a ) i+1 =MHA-VGAP(L((x p ) i )+(x p ) i ,i=0,...N

[0078] MHA-VGAP=MultiHead(VGAP(q),VGAP(k),VGAP(v))

[0079] VGAP(q)=q w:ap ·q ap +q w:mp ·q mp

[0080] Among them, (x p ) i is the feature vector with position encoding of the input of the i-th layer, L is the layer normalization, and MHA-VGAP is the multi-head attention calculation with visually guided adaptive pooling

[0081] (x p ) i+1 =MLP(L((x a ) i+1 ))+(x a ) i+1 ,i=0,1...N

[0082] The fusion of local features and global features

[0083] x m,n =w l ·x m +w g ·x n

[0084] Among them, w l ,w g The weights for adaptive dynamic allocation of local and global features respectively;

[0085] (3c) Innovatively introduce a dynamic bimodal attention module to achieve accurate representation of retinal blood vessels through a multi-dimensional feature enhancement strategy: Dynamic perception mechanism in spatial domain: A deformable convolution system based on residual learning, which dynamically adjusts the standard convolution grid p by designing an offset generation network o The sampling position is optimized through reinforcement learning, so that the model can adaptively capture the complex geometric topology of the vascular tree. It is combined with the feature non-local association enhancement unit: a spatial attention module based on non-local operations is constructed. By calculating the association weight between any two points in the feature map, the generated attention map can adaptively enhance the response of the vascular area. And channel dimension screening: a channel attention mechanism based on the squeeze-and-excite module is designed. Channel statistics are generated through global average pooling (GAP), and channel weights are generated through a two-layer fully connected network.

[0086] Its variable convolution calculation process:

[0087]

[0088] Among them, W k is the convolution kernel weight, X m,n is the input feature, p0 is the center position, p k is the regular network position, Δ pk is the offset for learning;

[0089] Its edge feature calculation process:

[0090]

[0091] Among them, σ is the Sigmoid function, G x , G y They are gradient operations in vertical and horizontal directions respectively;

[0092] The spatial attention weight calculation process:

[0093] W s =Sigmoid(Conv(GAP([X l ;X e ])))

[0094] Its channel attention weight calculation process:

[0095] W c =Sigmoid(FC(GAP(X l )))

[0096] Its adaptive feature fusion calculation process:

[0097]

[0098] (4) Introducing incremental learning technology to train the model. In the first stage, VGAT-Net-IL is trained using fundus image datasets such as STARE, so that the model can learn the feature representation of retinal blood vessel segmentation and basic lesions (such as microaneurysms), and output accurate blood vessel segmentation masks and lesion classification results. In the second stage, when new lesion data (such as exudate-related data) are introduced, continuous learning loss is used to optimize the model. The continuous learning loss consists of blood vessel segmentation loss, old lesion classification loss, new lesion classification loss, and mutual distillation loss;

[0099] Vessel segmentation loss (L seg ): It is used to measure the difference between the retinal vessel segmentation results predicted by the model and the actual vessel area.

[0100]

[0101] Where N is the total number of pixels, y i is the true label of pixel i, is the probability that the model predicts that pixel i belongs to a blood vessel.

[0102] Old lesion classification loss (L class-old ): Maintain the model's recognition accuracy for old lesions that have been learned.

[0103]

[0104] Where M is the number of old lesion samples, Cold is the number of old lesion categories, y jk Indicates whether sample j belongs to old lesion category k, p jk is the probability that the model predicts that sample j belongs to old lesion category k.

[0105] New lesion classification loss (L class-new ): Guide the model to learn the newly introduced lesion type features.

[0106]

[0107] Where N is the number of new lesion samples, C new is the number of new lesion categories, y lm Indicates whether the sample l belongs to the new lesion category m, p lm is the probability that the model predicts that sample l belongs to the new lesion category m.

[0108] Distillation loss (L md ): Calculate the joint probability distribution of new and old lesions based on the Bayesian rule, aiming to learn the structural and semantic dependencies between new and old lesions and establish associations between different lesions.

[0109]

[0110] Among them, C old For old lesion category, C new is the lesion category, p(c old ,c new ) is the prior joint probability of the new and old lesion categories, p(z|c old ,c new ) is the likelihood function of feature z given the new and old lesion categories, p(c old ) is the prior probability of the old lesion class, and p(z) is the probability of feature z.

[0111] (5) The trained VGAT-Net-IL network model is applied to the test image to segment the retinal blood vessels and detect lesions, and finally obtain high-precision retinal blood vessel segmentation result maps and lesion detection results.

Claims

1. A retinal vessel segmentation and lesion detection method based on incremental learning, characterized in that: The specific steps are as follows: (1) Implement a preprocessing strategy for the original retinal image. Use a region partitioning algorithm to segment the original retinal image into multiple sub-regions with specific spatial scales. For each sub-region image, apply data enhancement technology, including but not limited to rotation and flipping operations based on geometric transformation, to simulate retinal imaging conditions under different viewing angles. (2) Construct the VGAT-Net-IL network model, which adopts the encoding-decoding symmetric structure as the backbone architecture; (2a) First, the preprocessed data x in step (1) is p The data is input to the downsampling module to reduce the spatial resolution of the data and expand the receptive field, laying the foundation for deep feature extraction. Among them, x p is the preprocessed data, k1, k2 are the pooling kernel size and convolution kernel size respectively, s1, s2 are the step size of pooling and convolution operation respectively; (2b) The downsampled feature map x d The image is input into the adaptive receptive field module (VGA-Module). This module draws on the perception principle of the biological visual cortex and dynamically adjusts the scale and range of the receptive field according to the input image features, thereby synchronously extracting the local and global features of the retinal blood vessels. Its local features: x m =ResNet(x d )=CBblock n (...(CBblock2(CBblock1(x d )))...) Its global features, the multi-channel feature map x extracted in step (2a) d Perform spatial dimension division to generate non-overlapping image blocks of fixed size, and map them into feature vector sequences through linear projection. This encoding method embeds spatial position information into the feature dimension and preserves the relative position relationship between feature blocks. Its global characteristics: x p =[(x pj ·W)]+x ep Among them, x pj is the projected patch feature matrix, W is the weight matrix, x ep is the position encoding matrix; (x a ) i+1 =MHA-VGAP(L((x p ) i )+(x p ) i ,i=0,...N MHA-VGAP=MultiHead(VGAP(q),VGAP(k),VGAP(v)) VGAP(q)=q w:ap ·q ap +q w:mp ·q mp Among them, (x p ) i is the feature vector with position encoding, L is layer normalization, and MHA-VGAP is multi-head attention calculation with visually guided adaptive pooling; (x p ) i+1 =MLP(L((x a ) i+1 ))+(x a ) i+1 ,i=0,1...N The fusion of local features and global features: x m,n =w l ·x m +w g ·x n Among them, w l ,w g The weights for adaptive dynamic allocation of local and global features respectively; (2c) Innovatively introduce a dynamic bimodal attention module to achieve accurate representation of retinal blood vessels through a multi-dimensional feature enhancement strategy: Dynamic perception mechanism in spatial domain: A deformable convolution system based on residual learning, which dynamically adjusts the standard convolution grid p by designing an offset generation network o The sampling position is optimized through reinforcement learning, so that the model can adaptively capture the complex geometric topology of the vascular tree; it is combined with the feature non-local association enhancement unit: a spatial attention module based on non-local operation is constructed. By calculating the association weight between any two points in the feature map, the generated attention map can adaptively enhance the response of the vascular area; and channel dimension screening: a channel attention mechanism based on the squeeze-and-excite module is designed. Channel statistics are generated through global average pooling (GAP), and channel weights are generated through a two-layer fully connected network. Its variable convolution calculation: Among them, W k is the convolution kernel weight, X m,n is the input feature, p0 is the center position, p k is the regular network position, Δ pk is a learnable offset; Its edge feature calculation: Among them, σ is the Sigmoid function, G x , G y They are gradient operations in vertical and horizontal directions respectively; Its spatial attention weight calculation: W s =Sigmoid(Conv(GAP([X l ;X e ]))) Its channel attention weight calculation: W c =Sigmoid(FC(GAP(X l ))) Its adaptive feature fusion calculation: (3) Introduce incremental learning technology to train the model. Basic training stage: Use fundus image datasets such as STARE to train the model to learn the feature representation of retinal blood vessel segmentation and basic lesions (such as microaneurysms), and output accurate blood vessel segmentation masks and lesion classification results. Incremental optimization stage: When introducing new lesion data (such as exudate-related data), adopt a multi-loss function joint optimization strategy: blood vessel segmentation loss, old lesion classification loss, new lesion classification loss and mutual distillation loss; (4) The trained VGAT-Net-IL network model is applied to the test image to perform retinal vessel segmentation and lesion detection tasks, and finally obtain high-precision retinal vessel segmentation result images and lesion detection results.

2. The retinal blood vessel segmentation and lesion detection method according to claim 1, characterized in that Step (2b): Input the downsampled features into the adaptive receptive field module, which extracts the local features x m and global features x n After that, it is transmitted to the fusion unit together with the input features. In the fusion unit, each feature is first normalized, and then the adaptive weight generation module calculates the fusion weight W l , W g ,Finally, each feature is fused according to this weight to obtain the adaptive fusion feature; Its weight calculation: in, are the results of input feature normalization, f m , f n is a linearly changing function.

3. The retinal blood vessel segmentation and lesion detection method according to claim 1, characterized in that Step (2c): Input the fused features into the dynamic bimodal attention module. In this module, variable convolution is used to perform X on the fused features. m,n Further feature extraction, at the same time, the edge detection algorithm is used to obtain the blood vessel edge feature X e In addition, a channel detection operation is performed to analyze the channel information of the fused features.

4. The retinal blood vessel segmentation and lesion detection method according to claim 1, characterized in that Step (2c): Obtain the variable convolution feature X l With edge detection feature X e Finally, the spatial attention weight map W is realized by modeling the correlation between different positions of the feature map s Derivation of ; By performing global average pooling and full connection layer calculation on the channel information, the channel attention weight vector W is obtained c . W s =Sigmoid(Conv(GAP([X l ;X e ]))) W c =Sigmoid(FC(GAP(X l )))。 5. The retinal blood vessel segmentation and lesion detection method according to claim 1, characterized in that Step (2c): Based on the calculated spatial attention weight map W s And the channel attention weight vector W c , for variable convolution feature X l and edge detection feature X e Perform weighted fusion processing to obtain the final fusion features.

6. The retinal blood vessel segmentation and lesion detection method according to claim 1, characterized in that In step (3), in order to solve the multi-task optimization problem faced by the model in the incremental learning process, that is, to ensure the high accuracy of retinal blood vessel segmentation while taking into account the maintenance of the ability to recognize old lesions, the learning of new lesion features, and the mining of semantic associations between lesions, a multi-loss function collaborative optimization strategy is adopted: blood vessel segmentation loss, old lesion classification loss, new lesion classification loss and mutual distillation loss, to collaboratively optimize the model from multiple dimensions. Vessel segmentation loss (L seg ): Where N is the total number of pixels, y i is the true label of pixel i, is the probability that the model predicts that pixel i belongs to a blood vessel; Old lesion classification loss (L class-old ): Where M is the number of old lesion samples, C old is the number of old lesion categories, y jk Indicates whether sample j belongs to old lesion category k, p jk is the probability that the model predicts that sample j belongs to old lesion category k; New lesion classification loss (L class-new ): Where N is the number of new lesion samples, C new is the number of new lesion categories, y lm Indicates whether the sample l belongs to the new lesion category m, p lm is the probability that the model predicts that sample l belongs to the new lesion category m; Distillation loss (L md ): Among them, C old For old lesion category, C new is the lesion category, p(c old ,c new ) is the prior joint probability of the new and old lesion categories, p(z|c old ,c new ) is the likelihood function of feature z given the new and old lesion categories, p(c old ) is the prior probability of the old lesion class, and p(z) is the probability of feature z.

Images