Retinal vessel segmentation method based on N-LOSAM2 network

Abstract

The invention provides a retinal vessel segmentation method based on an N-LOSAM2 network, and the method comprises the steps: firstly inputting a retinal fundus image, carrying out the preprocessing of a low-quality image through cofe-Net, cutting the image into image blocks through a sliding window, enabling the adjacent image blocks to have an overlapping part, enabling the obtained image blocks to form an image sequence, and carrying out the segmentation of the retinal fundus image. Image blocks are input into a learnable prompt layer to learn medical priori knowledge such as shapes and distribution of retinal blood vessels, a CLTransform module is combined to reinforce recognition of small blood vessel features, an image encoder is used for encoding the embedded image block features, then the encoded image block features are transmitted to a mask decoder, and the mask decoder decodes the encoded image block features. A user can provide prompts in a mode of clicking, frame selection or mask providing, the model predicts the segmentation mask of the retinal blood vessel frame by frame according to the prompts and a memory mechanism, the segmentation of the current frame is guided by utilizing the segmentation result of the previous frame in a memory library, and finally, the accurate blood vessel segmentation mask is output through a task head. Therefore, efficient and accurate retinal vessel segmentation is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a retinal blood vessel segmentation method based on an N-LOSAM2 network. Background Art

[0002] Retinal vessel segmentation is a crucial topic in the field of medical image segmentation. It can be used for early diagnosis and treatment of a variety of ophthalmic diseases, such as diabetic retinopathy, hypertensive retinopathy, age-related macular degeneration and other fundus diseases. Studies have shown that retinal blood vessels in diabetic patients often have abnormalities, and retinal vessel segmentation technology can assist doctors in accurately determining the location and extent of these abnormalities. Traditional methods usually use unsupervised learning techniques (such as vessel tracking, matched filtering, and level sets) to build vessel segmentation models. However, since these methods mainly rely on local features and morphological information of the image, they cannot accurately segment retinal vessels, especially in areas where vessels cross, end, or have low contrast. This is because they are difficult to capture the global structure and contextual information of the vessels. In addition, unsupervised learning methods are easily affected by the individual characteristics of the data and cannot be widely used, and the segmentation accuracy is difficult to improve. With the development of medical image processing technology, researchers have proposed a large number of retinal vascular segmentation methods. According to the degree of automation of the segmentation method, these methods can be divided into two categories: traditional segmentation methods and deep learning-based segmentation methods. In terms of traditional segmentation, Khan et al. proposed a method to divide retinal vessels into large vessels and small vessels in "Width-wise vessel bifurcation for improved retinal vessel segmentation.", and then enhanced and detected each group of vessels separately. Yang et al. proposed a method based on Frangi multi-scale level sets to segment retinal vessels in fundus images in "Frangi based multi-scale level sets for retinal vascular segmentation." The vascular structure is enhanced by Frangi filters, and the local optimal scale is obtained at the same time, which is used as the input of the proposed level set model. Orujov et al. proposed a method based on Mamdani (Type-2) fuzzy rules in "Fuzzybased image edge detection algorithm for blood vessel detection in retinal images." They used the green channel of the fundus image, enhanced the contrast through Contrast Limited Adaptive Histogram Equalization (CLAHE), removed the background with median filtering, and finally used fuzzy rules for edge detection. On the other hand, deep learning-based segmentation methods automatically learn prior knowledge and deep features through supervised learning techniques, thereby improving the accuracy of retinal vessel segmentation. In recent years, deep learning-based retinal vessel segmentation methods mainly use network models based on CNN, U-Net, GAN, etc. Wu et al. proposed a CNN-based network following network (NFN+) model in "A novel network followed network for retinal vessel segmentation". They designed a multi-scale backbone network containing two identical networks, and enhanced feature transfer through internal and inter-network jump connections, solving the problems of insufficient recognition of small blood vessels, limited data volume, and structured prediction in traditional methods. However, although the fuzzy details of the blood vessels can be found from the segmented blood vessel images themselves by cascading multiple sub-networks, they cannot adapt well to complex blood vessel trees due to the limitations of the baseline network itself. Matloob et al. proposed a lightweight multi-path bidirectional skip connection based multipath CNN for retinal blood vessel segmentation, which consists of a bottleneck layer, a multi-path block, three encoder blocks, and three decoder blocks, which are arranged in a systematic way to solve the gradient flow problem and help preserve fine-grained details. However, the performance of small blood vessel segmentation is affected by scaling on high-resolution images. Wei et al. proposed a Genetic U-Net network based on U-Net in Genetic U-Net: automatically designed deep networks for retinal vessel segmentation using a genetic algorithm, which uses a genetic algorithm to optimize the U-shaped network structure and reduce model parameters and computational burden. However, its performance is limited by the number of training samples and the quality of annotations. In practice, it is costly to obtain accurately annotated retinal vascular images, which limits the application of supervised learning. Tang et al. proposed a GAN-based network in The joint learning of multi-resolution feature for multi-class retinal vessel segmentation, which fuses cross-scale features through a global-local dual-branch structure and multi-kernel size convolution, improving the problems of arteriovenous confusion and discontinuity of vascular edges.However, due to resource constraints, the lack of experiments on the original image resolution will affect the actual performance and application of the method. Zhou et al. proposed a large-model-based automated retinal vascular morphology quantification tool in "AutoMorph: automated retinal vascular morphology quantification via a deep learning pipeline", which includes four functional modules, namely image preprocessing, image quality rating, structure segmentation (including vascular segmentation, artery / vein segmentation and optic disc segmentation) and vascular morphology feature measurement. In the retinal vascular segmentation task, AutoMorph uses an adversarial segmentation network to improve the segmentation performance of vascular structures, and uses an information fusion network to optimize arteriovenous segmentation. Qiu et al. introduced a learnable hint layer based on the SAM large model framework in "Learnable ophthalmology sam", which adapts to multimodal segmentation tasks of ophthalmic images, such as vascular and lesion segmentation, by learning medical prior knowledge. It improves generalization ability and simplifies the training process by training the hint layer and task head once. Although LOSAM performs well in multi-task segmentation, its performance is still limited when dealing with small targets, especially when the contrast between the target and the background is low. On the other hand, when the retinal fundus image has factors such as image blur and artifacts that make the edge features of blood vessels unclear, it is difficult for the algorithm to accurately identify them. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention proposes a retinal blood vessel segmentation method based on the N-LOSAM2 network, which optimizes and improves LOSAM from two aspects: improving the low-quality fundus images and small blood vessel recognition problems, and solves the problem of unsatisfactory performance of traditional LOSAM when processing small targets or low-quality images in the prior art. To achieve the above object, the present invention adopts the following technical solutions: A retinal blood vessel segmentation method based on the N-LOSAM2 network, wherein the N-LOSAM2 network replaces the original SAM network architecture in LOSAM with Segment Anything Model 2, and introduces a cofe-Net network and a CLTransformer module. The method comprises the following steps: Step 1: Use the cofe-Net module to preprocess the retinal fundus image to remove degradation factors; Step 2: Crop the image obtained in step 1 into image blocks through a sliding window, and compose the image blocks into an image sequence; Step 3: Input the image sequence in step 2 into the learnable hint layer through image block embedding, and refine the image features through the CLTransformer module; Step 4: Input the refined image features into the image encoder for encoding to obtain an encoded image sequence; pass the encoded image sequence to the mask decoder, which predicts the segmentation mask of the retinal vessels in the image sequence frame by frame based on user prompts, and uses the segmentation results of the previous frames in the memory bank to guide the segmentation of the current frame; Step 5: Output the blood vessel segmentation mask in the image sequence through the task head to obtain the retinal blood vessel segmentation of the retinal fundus image. Preferably, in step 1, the process of preprocessing the retinal fundus image using the cofe-Net module is as follows: 1) Input low-quality retinal fundus images and standardize the input data; 2) Identify and focus on low-quality areas in the image through the LQA module and remove low-quality factors; 3) Predict the segmentation mask of the retinal structure in the retinal fundus image through the RSA module, and retain the clinical features in the retinal fundus image; 4) The image correction module fuses the information of the LQA module and the RSA module as well as the potential features of the image to output the final corrected image. Preferably, in step 2, the process of cropping the image into image blocks through a sliding window and composing the image blocks into an image sequence is as follows: the sliding window starts from the upper left corner of the image and moves from left to right and from top to bottom with a step length d, and each movement captures a part of the image, which becomes a frame in the sequence; an overlapping area is designed between every two adjacent frames, and the overlapping ratio is r. The captured image sequence can be expressed as: S = {W 1 ,W 2 ,…,W n}; where W is a custom sliding window and S is a captured image sequence. Preferably, in step three, the attention part of the CLTransformer module includes layer normalization and cross-layer attention, and a position feedforward network is connected after the cross-layer attention. The position feedforward network layer consists of two fully connected feedforward neural networks and a residual connection. The residual connection formula is: X out =(ReLU(X in ω 1 +b 1 ))ω 2 +b 2 +X in Among them, X in and Xout Respectively represent the input features and output features of the position feedforward network layer, ω 1 and b 1 ,ω 2 and b 2 They represent the learning parameters of two fully connected feedforward neural networks respectively. Preferably, in step 4, the captured image sequence S is input into the image encoder E for encoding to obtain the encoded feature sequence F=E(S); the mask decoder predicts the segmentation mask Mi=Predict(Fi,P,M) of the retinal blood vessels frame by frame according to the user prompt; Where Fi is the i-th encoded feature frame, M is the memory mechanism, and P is the prompt provided by the user on any frame; If the segmentation is wrong, refine the segmentation result according to the further prompt P′: Mi′=Refine(Mi,P′,R); Wherein R is the segmentation result of the previous frame; the segmentation accuracy is evaluated by the evaluation index G: G=Evaluate(Mi,Mi′), and the evaluation index G includes Dice coefficient, Jaccard index, sensitivity, specificity, and precision. Preferably, the process of predicting the segmentation mask image of the retinal structure in the retinal fundus image by the RSA module is: in and denote the features extracted from the RSA module and the nth layer of the image correction network, respectively. Concat is the concatenation operator. T n S is a learnable nonlinear transformation filter, S L is the output segmentation mask map. Preferably, the process of fusing the information of the LQA module and the RSA module and the potential features of the image through the image correction module to output the final corrected image is: Where I is the input image, W G are the learnable parameters of the network G, is the rectified image; Loss function: Where L m is the loss of the LQA module, and They are the image reconstruction losses at different scales, L v is the MSE loss of the RSA module, s represents the scale index, and λ p ,λ c ,λ v are the weights of different loss terms. Beneficial effects: 1. The present invention introduces the cofe-Net module to optimize the processing of low-quality fundus images. Through the collaborative work of LQA, RSA and image correction modules, the image quality is significantly improved, providing clearer image features for subsequent vascular segmentation. 2. The present invention enhances the model's ability to identify small blood vessels by introducing the CLTransformer module. This module effectively integrates low-level details and high-level semantic information through the combination of CAM and position feedforward network, thereby improving the accuracy of small blood vessel segmentation. 3. The present invention replaces the original SAM network architecture in LOSAM with SAM2. Through streaming processing, the video is processed frame by frame, and the memory module is interactively refined, aiming to further improve the speed of the model and the accuracy of retinal blood vessel segmentation. BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 This is a diagram of the N-LOSAM2 network structure of an embodiment of the present invention; Figure 2 This is a LOSAM2 network structure diagram of an embodiment of the present invention; Figure 3 A diagram of the SAM2 network structure of an embodiment of the present invention; Figure 4 A cofe-Net framework diagram of an embodiment of the present invention; Figure 5 It is a CLTransformer module of an embodiment of the present invention. DETAILED DESCRIPTION The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. like Figure 1 As shown, the N-LOSAM2 network structure of the present invention integrates the improved large-model-based LOSAM, cofe-Net and CLTransformer modules, uses cofe-Net to process low-quality fundus images, introduces the CLTransformer module after the learnable prompt layer, and uses SAM2 to replace the original SAM network architecture in LOSAM. like Figure 2 As shown in FIG. 1 , the improved LOSAM in the present invention has significant advantages over the traditional LOSAM, and the improved LOSAM2 is obtained by replacing the SAM framework in the original LOSAM model with the SAM2 framework. This design not only improves the processing speed of the model, but also enhances the accuracy and robustness of the model in the retinal vessel segmentation task. like Figure 3 The SAM2 framework shown uses more advanced data processing and feature extraction techniques, making LOSAM2 more efficient in processing retinal images. SAM2's streaming processing method allows the model to process video data frame by frame, which is particularly important for real-time or near-real-time clinical applications. Secondly, a memory module is introduced, which can store and utilize past segmentation results to assist the segmentation task of the current frame. This memory mechanism is particularly helpful for tracking blood vessels in video sequences, and can maintain the continuity of segmentation even when the blood vessels are temporarily occluded or exit the screen and then reappear. The MAE pre-trained Hiera encoder is used in the SAM2 framework, which can generate multi-scale features to help LOSAM2 better understand the details and contextual information in the image, thereby improving the accuracy of segmentation. LOSAM2 supports interactive segmentation, allowing users to provide prompts on the image by clicking, selecting, or providing masks, and the model generates segmentation masks based on these prompts. This interactive design enables the model to better adapt to the user's specific needs and make real-time corrections during the segmentation process. The method of the present invention specifically comprises the following steps: Step 1: Use the cofe-Net module to preprocess the retinal fundus image to remove degradation factors; Step 2: Crop the image obtained in step 1 into image blocks through a sliding window, and compose the image blocks into an image sequence; Step 3: Input the image sequence in step 2 into the learnable hint layer through image block embedding, and refine the image features through the CLTransformer module; Step 4: Input the refined image features into the image encoder for encoding to obtain the encoded image sequence; and pass it to the mask decoder, which predicts the segmentation mask of the retinal vessels in the image sequence frame by frame according to the user prompt, and uses the segmentation results of the previous frames in the memory bank to guide the segmentation of the current frame; Step 5: Output the blood vessel segmentation mask in the image sequence through the task head to obtain the retinal blood vessel segmentation of the retinal fundus image. In step 1, the cofe-Net network structure is as follows Figure 4 As shown in Figure 1, the input retinal fundus image is of low quality (may contain degradation factors such as uneven illumination, image blur and artifacts) Where W is the image width, H is the image height, and C is the number of channels; the input data is standardized: where μ and σ are the mean and standard deviation of the training data, respectively. Through the LQA (Low Quality Activation) module, the network is able to identify and focus on low-quality areas in the image, such as artifacts; formula: m = σ(M(I,W M ), where I is the input image, W M are the parameters of the LQA module, σ is the nonlinear activation function, and m is the output artifact attention map. The RSA (Retinal Structure Activation) module guides the image reconstruction process by predicting segmentation masks of major retinal structures, ensuring that important clinical features such as blood vessels, optic disc and cup areas are preserved during the correction process. The process is: in and denote the features extracted from the RSA module and the nth layer of the image correction network, respectively. Concat is the concatenation operator. T n S is a learnable nonlinear transformation filter, S L is the output segmentation mask map. The image correction module is used to fuse the information of the LQA and RSA modules and the potential features of the image to produce the final corrected image. The process is as follows: input a low-quality retinal fundus image and feature maps from the LQA and RSA modules. The formula is: Where I is the input image, W G are the learnable parameters of the network G, is the rectified image; The loss function is: Where L m is the loss of the LQA module, and They are the image reconstruction losses at different scales, L v is the MSE loss of the RSA module, s represents the scale index, and λ p ,λ c ,λ v are the weights of different loss terms. The corrected high-quality retinal fundus images have improved image quality while retaining important clinical features such as vascular structure and pathological characteristics. In step 2, if Figure 1 As shown, let I be a complete retinal fundus image, W be a custom sliding window, and S be a captured image sequence. The sliding window starts from the upper left corner of the image and moves from left to right and from top to bottom with a certain step length d. Each movement captures a part of the image, which becomes a frame in the sequence. An overlapping area is designed between every two adjacent frames, and the overlapping ratio is r. Therefore, the captured image sequence can be expressed as: S = {W1, W2, ..., Wn}; where Wi is the i-th captured image frame. In step 3, the CLTransformer module structure is as follows Figure 5 As shown, the input of the CLTransformer module includes: Encoder features: semantic information from deeper layers of the network. Decoder features: feature information from shallower layers of the network. The output is refined features: features processed by CLTransformer for further blood vessel segmentation; The residual connection part promotes the direct flow of information in the module, avoiding the problem of gradient disappearance and information loss in deep networks. It can be expressed as: X out =(ReLU(X in ω 1 +b 1 ))ω 2 +b 2 +X in Among them, X in and X out Respectively represent the input features and output features of the position feedforward network layer, ω 1 and b 1 ,ω 2 and b 2 They represent the learning parameters of two fully connected feedforward neural networks respectively. In step 4, if Figure 3 As shown in the SAM2 network architecture, let E be an image encoder, input the captured image sequence S into the image encoder for encoding, and obtain the encoded feature sequence F = E(S); Let P be the hint provided by the user on any frame of the video, M be the memory mechanism, and the mask decoder predicts the segmentation mask Mi = Predict(Fi, P, M) of the retinal vessels frame by frame based on the hint provided by the user. Where Fi is the i-th encoded feature frame. Refine the segmentation result: Let R be the segmentation result of the previous frame. If the segmentation is wrong, refine the segmentation result according to the further prompt P′: Mi′=Refine(Mi,P′,R); Segmentation accuracy evaluation: Let G be the segmentation accuracy evaluation index of retinal blood vessels. The Dice coefficient, Jaccard index, sensitivity, specificity, precision and other indicators can be used to evaluate the accuracy of segmentation, G = Evaluate (Mi, Mi′). In summary, the improved LOSAM2 in the present invention replaces the SAM framework in the original LOSAM model with the SAM2 framework. This design not only improves the processing speed of the model, but also enhances the accuracy and robustness of the model in the retinal blood vessel segmentation task. The SAM2 framework adopts more advanced data processing and feature extraction techniques, making LOSAM2 more efficient in processing retinal images. The streaming processing method of SAM2 allows the model to process video data frame by frame, which is particularly important for real-time or near-real-time clinical applications. Secondly, a memory module is introduced, which can store and use past segmentation results to assist the segmentation task of the current frame. This memory mechanism is particularly helpful for tracking blood vessels in video sequences, and can maintain the continuity of segmentation even when the blood vessels are temporarily occluded or exit the screen and reappear. The MAE pre-trained Hiera encoder is used in the SAM2 framework, which can generate multi-scale features to help LOSAM2 better understand the details and contextual information in the image, thereby improving the accuracy of segmentation. LOSAM2 supports interactive segmentation, allowing users to provide prompts on the image by clicking, boxing or providing masks, and the model generates segmentation masks based on these prompts. This interactive design allows the model to better adapt to the user's specific needs and make real-time corrections during the segmentation process. The purpose of using cofe-Net to process low-quality fundus images is to improve the quality of retinal vascular images. cofe-Net contains three modules, namely LQA (Low-Quality Activation Module), RSA (Retinal Structure Activation Module) and image correction module. The role of the LQA module is to identify and activate the blurred, low-contrast and artifact parts in the low-quality retinal vascular images, mark these parts, generate attention maps to guide the image correction process and retain the true retinal structure. The role of the RSA module is to activate and protect the key retinal vascular structures in the fundus images to ensure that these anatomical features that are critical to clinical diagnosis are retained during the image correction process. The RSA module continuously injects retinal structural information into the correction network through the feature map connection at each scale. The role of the image correction module is to fuse the information provided by the LQA and RSA modules, gradually enhance the image quality through the encoding-decoding network, while retaining the clinically important structures and pathological features in the fundus images, and generate low-resolution corrected images. Finally, the low-resolution correction result is upsampled to the original resolution and then input into the image correction module together with the original image for high-resolution correction to generate the final high-quality corrected image. The CLTransformer module is introduced after the learnable hint layer to improve the model's ability to identify and segment small blood vessels. The CLTransformer module consists of two parts: the attention part and the feedforward network. In the attention part, layer normalization (LayerNorm) and cross-layer attention (CAM) work together to enhance the expressiveness of vascular features at each layer and determine the logical relationship of image context information. CAM strengthens the model's fusion of features at different levels, so that low-level detail information and high-level semantic information, such as the connection relationship between vascular branches, can be effectively integrated, thereby improving the accuracy of the segmentation task. Since CAM mainly establishes dependencies between layers, it may lack information attention in the position dimension. In order to compensate for this and further refine the topological geometric position of the blood vessels, a position feedforward network is connected after the cross-layer attention structure to construct more accurate position semantic information, thereby improving the accuracy of small blood vessel segmentation. The position feedforward network layer consists of two fully connected feedforward neural networks and residual connections. The fully connected feedforward neural network part acts as a nonlinear transformation to help the module learn more complex and advanced retinal vascular features. The residual connection part promotes the direct flow of information in the module, effectively avoiding the common gradient vanishing and information loss problems in deep networks. Through this structural design, CLTransformer not only improves the segmentation performance of large blood vessels, but also improves the ability to identify and segment small blood vessel structures. Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if these modifications and variations of the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A retinal blood vessel segmentation method based on N-LOSAM2 network, characterized in that: The N-LOSAM2 network replaces the original SAM network architecture in LOSAM with Segment Anything Model 2, and introduces the cofe-Net network and CLTransformer module. The method includes the following steps: Step 1: Use the cofe-Net module to preprocess the retinal fundus image to remove degradation factors; Step 2: Crop the image obtained in step 1 into image blocks through a sliding window, and compose the image blocks into an image sequence; Step 3: Input the image sequence in step 2 into the learnable hint layer through image block embedding, and refine the image features through the CLTransformer module; Step 4: Input the refined image features into the image encoder for encoding to obtain an encoded image sequence; pass the encoded image sequence to the mask decoder, which predicts the segmentation mask of the retinal vessels in the image sequence frame by frame based on user prompts, and uses the segmentation results of the previous frames in the memory bank to guide the segmentation of the current frame; Step 5: Output the blood vessel segmentation mask in the image sequence through the task head to obtain the retinal blood vessel segmentation of the retinal fundus image.

2. The retinal blood vessel segmentation method based on N-LOSAM2 network according to claim 1, characterized in that: In step 1, the process of preprocessing the retinal fundus image using the cofe-Net module is as follows: 1) Input low-quality retinal fundus images and standardize the input data; 2) Identify and focus on low-quality areas in the image through the LQA module and remove low-quality factors; 3) Predict the segmentation mask of the retinal structure in the retinal fundus image through the RSA module, and retain the clinical features in the retinal fundus image; 4) The image correction module fuses the information of the LQA module and the RSA module as well as the potential features of the image to output the final corrected image.

3. The retinal blood vessel segmentation method based on N-LOSAM2 network according to claim 1, characterized in that: In step 2, the image is cut into image blocks through a sliding window, and the process of composing an image sequence with the image blocks is as follows: the sliding window starts from the upper left corner of the image and moves from left to right and from top to bottom with a step length d. Each movement captures a part of the image, which becomes a frame in the sequence; an overlapping area is designed between every two adjacent frames, and the overlapping ratio is r. The captured image sequence can be expressed as: S = {W1, W2, ..., W n }; Where W is a custom sliding window and S is the captured image sequence.

4. The retinal blood vessel segmentation method based on N-LOSAM2 network according to claim 1, characterized in that: In step 3, the attention part of the CLTransformer module includes layer normalization and cross-layer attention. After the cross-layer attention, the position feedforward network is connected. The position feedforward network layer consists of two fully connected feedforward neural networks and residual connections. The residual connection formula is: X out =(ReLU(X in ω1+b1))ω2+b2+X in Among them, X in and X out They represent the input features and output features of the position feedforward network layer respectively, ω1 and b1, ω2 and b2 represent the learning parameters of two fully connected feedforward neural networks respectively.

5. The retinal blood vessel segmentation method based on N-LOSAM2 network according to claim 1, characterized in that: In step 4, the captured image sequence S is input into the image encoder E for encoding to obtain the encoded feature sequence F = E(S); the mask decoder predicts the segmentation mask Mi = Predict(Fi, P, M) of the retinal blood vessels frame by frame according to the user prompt; Where Fi is the i-th encoded feature frame, M is the memory mechanism, and P is the prompt provided by the user on any frame; If the segmentation is wrong, refine the segmentation result according to the further prompt P′: Mi′=Refine(Mi,P′,R); Wherein R is the segmentation result of the previous frame; the segmentation accuracy is evaluated by the evaluation index G: G=Evaluate(Mi,Mi′), and the evaluation index G includes Dice coefficient, Jaccard index, sensitivity, specificity, and precision.

6. The retinal blood vessel segmentation method based on N-LOSAM2 network according to claim 2, characterized in that: The process of standardizing the input data is as follows: Where μ and σ are the mean and standard deviation of the training data, respectively. is the input low-quality image, where W is the image width, H is the image height, and C is the number of channels.

7. The retinal blood vessel segmentation method based on N-LOSAM2 network according to claim 2, characterized in that: The process of predicting the segmentation mask image of the retinal structure in the retinal fundus image by the RSA module is as follows: in and Represent the features extracted from the nth layer of the RSA module and the image correction module respectively, Concat is the connection operator, T n S is a learnable nonlinear transformation filter, S L is the output segmentation mask map.

8. The retinal blood vessel segmentation method based on N-LOSAM2 network according to claim 2, characterized in that: The process of fusing the information of the LQA module and the RSA module and the potential features of the image through the image correction module to output the final corrected image is as follows: Where I is the input image, W G are the learnable parameters of the image correction module G, is the rectified image; loss function: Where L m is the loss of the LQA module, and They are the image reconstruction losses at different scales, L v is the MSE loss of the RSA module, s represents the scale index, and λ p ,λ c ,λ v are the weights of different loss terms.

Images