Fundus image segmentation method based on uncertainty and shallow feature enhanced attention network

Abstract

The present invention discloses a fundus image segmentation method based on an uncertainty and shallow feature enhanced attention network, which comprises: selecting fundus retinal image data as a training set and a test set; preprocessing the fundus retinal images in the training set; constructing a fully convolutional neural network model on Pytorch; and segmenting the test set using the trained fully convolutional neural network model to obtain a final segmentation result. The present invention provides a neural network model for automatically segmenting retinal blood vessels from fundus images, which combines an attention mechanism to expand the receptive field so that the model can better understand the local details and global context in the image, and improves the learning ability of the model by fusing deep features in the neural network to guide the expression of shallow features. By quantifying image uncertainty, the adverse effects of the model's overconfidence on the segmentation results are reduced, thereby improving the accuracy of retinal blood vessel segmentation.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a fundus image segmentation method based on uncertainty and shallow feature enhanced attention network. Background Art

[0002] The number of patients suffering from blinding eye diseases such as glaucoma, diabetic retinopathy (DR) and age-related macular degeneration (AMD) is large, affecting more than 418 million people worldwide. The early development of these eye diseases is usually unnoticeable to patients, and clinical early screening is achieved by detecting retinal structures in the fundus of patients because they are highly correlated with changes in retinal morphology. Previously, this work was done manually by experienced experts, but annotating large datasets is time-consuming and expensive. In addition, there are few experts who can complete this work independently, making it a challenging task. Therefore, it is still necessary to develop an efficient and automatic retinal vessel segmentation method.

[0003] In recent years, deep learning has revolutionized the field of medical image analysis, achieving outstanding results in a wide range of tasks, including image segmentation, classification, and detection. Convolutional neural networks (CNNs) excel in learning hierarchical features directly from raw image data, automatically extracting complex patterns and structures without the need for handcrafted features. A variety of deep learning-based methods have been proposed for vessel segmentation in fundus images, utilizing different network architectures and training strategies to improve segmentation accuracy. These methods typically involve fully convolutional networks (FCNs), U-Net architectures, and more recently, attention mechanisms and multi-scale feature integration to capture global context and detail information. Despite significant progress, challenges such as variations in vessel thickness, low contrast between vessels and background, and the presence of pathological areas remain obstacles to achieving robust segmentation performance.

[0004] In the study of fundus vessel segmentation, many research problems still exist. For example, (1) Vascular details: The vascular structure in fundus images is complex, including the main trunk and numerous small branches. The segmentation of small branches is particularly difficult because they may be almost invisible in low-resolution images. (2) Vascular crossing and overlap: The blood vessels may cross or overlap in the image, which makes it difficult for the segmentation algorithm to distinguish the boundaries of different blood vessels. (3) Low contrast: The contrast between blood vessels and the background in fundus images is usually low, especially under pathological conditions, which increases the difficulty of segmentation. (4) Foreground-background imbalance: The area of ​​blood vessels in the image is usually very small, and the number of pixels in the foreground (blood vessels) is seriously insufficient compared to the background. This class imbalance will affect the training effect of the segmentation algorithm. In addition, certain diseases (such as diabetic retinopathy) can cause abnormal spots or hemorrhages in fundus images. These lesions may interfere with the normal operation of the vascular segmentation algorithm. Summary of the Invention

[0005] In response to the defects in the existing technology, the present invention provides a fundus image segmentation method based on uncertainty and shallow feature enhanced attention network, which reduces the adverse effects of uncertainty by modeling uncertainty and combining it with deep learning methods; it also uses an attention module to focus on segmentation details, and at the same time uses deep feature information to enhance shallow feature information to improve the final segmentation effect.

[0006] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0007] A fundus image segmentation method based on uncertainty and shallow feature enhanced attention network includes the following steps:

[0008] Step 1): Select fundus retinal image data as training set and test set, each fundus retinal image sample includes the original color fundus retinal image and the corresponding blood vessel segmentation label;

[0009] Step 2): Preprocess the fundus retinal images in the training set;

[0010] Step 3): Build a neural network model on Pytorch;

[0011] Step 4): Use the trained neural network model to segment the test set and obtain the final segmentation result.

[0012] As a possible implementation, further, step 2) specifically includes the following steps:

[0013] Each fundus retinal image was grayscaled and normalized, and then the overall contrast was adjusted using contrast-limited adaptive histogram equalization (CLAHE) and gamma correction with a gamma value set to 1.5;

[0014] Each fundus retinal image and the corresponding retinal vessel segmentation labels are divided into smaller blocks or sub-images.

[0015] As a possible implementation, each fundus retinal image and the corresponding retinal vessel segmentation label are further divided into smaller blocks or sub-images, specifically:

[0016] The training set image data is randomly cropped into 100,000 images with a width and height of 48*48.

[0017] As a possible implementation, further, step 3) specifically includes the following steps:

[0018] Step 3.1): Input the original image into the network and obtain the features through a convolution, BN, and ReLU operation. Figure 1 , and then undergo three consecutive downsampling and convolution, BN, and ReLU operations to obtain feature maps 2, 3, and 4, a total of four feature maps;

[0019] Step 3.2): The features Figure 1 ~3 inputs the attention module to obtain attention feature maps a1, a2, a3;

[0020] Step 3.3): Input the attention feature map a3 and feature map 4 into the shallow feature enhancement module, and output feature map A3; input feature map A3 and attention feature map a2 into the shallow feature enhancement module, and output feature map A2; input feature map A2 and attention feature map a1 into the shallow feature enhancement module, and output feature map A1;

[0021] Step 3.4): After upsampling feature map 4, it is spliced ​​with feature map A3 on the channel, and then subjected to convolution, BN, and ReLU operations to obtain feature map U3; after upsampling feature map U3, it is spliced ​​with feature map A2 on the channel, and then subjected to convolution, BN, and ReLU operations to obtain feature map U2; after upsampling feature map U2, it is spliced ​​with feature map A1 on the channel, and then subjected to convolution, BN, and ReLU operations to obtain feature map U1;

[0022] Step 3.5): Perform upsampling and convolution operations on feature maps U2 and U3 by 2 and 4 times respectively to make their shapes the same as feature map U1. Then, concatenate feature maps U1, U2, and U3 on the channel to obtain feature map C. Input feature map C into the attention module to obtain feature map C1. Finally, perform residual addition operation on feature map C and feature map C1 to obtain the coarse segmentation map Seg.

[0023] Step 3.6): Perform uncertainty modeling on the coarse segmentation map Seg to obtain the evidence feature map E and the uncertainty feature map F. Generate a reliable mask M by combining the coarse segmentation image Seg with the evidence feature map E and the uncertainty feature map F. Finally, concatenate M and the coarse segmentation map Seg on the channel and perform a 1×1 convolution to output the final refined segmentation result.

[0024] As a possible implementation, further, the attention module is composed of a channel attention module and a spatial attention module;

[0025] The channel attention module is used to pass the input feature map through two parallel MaxPool layers and AvgPool layers to transform the feature map from C×H×W to C×1×1, and then process it through the Share MLP module; wherein the Share MLP module is used to compress the number of channels to 1 / r (reduction rate) times the original number of channels, and then expand it to the original number of channels, and obtain two activated results through the ReLU activation function;

[0026] Next, the two output results are added element by element, and then a sigmoid activation function is used to obtain the output of Channel Attention. This output is then multiplied by the original image to return to the size of C×H×W;

[0027] The spatial attention module is used to perform maximum pooling and average pooling operations on the output results of the channel attention module to obtain two 1×H×W feature maps, then concatenate the two feature maps through a Concat operation, and then obtain four feature maps through four dilated convolutions with different dilation rates. The four feature maps are then concatenated through a Concat operation, and then a sigmoid operation is performed to obtain a feature map of spatial attention. Finally, the output result is multiplied by the original image to return to the size of C×H×W.

[0028] As a possible implementation method, further, the shallow feature enhancement module is used to perform a Concat operation on the shallow features and the deep features after convolution and BN, and then input the spliced ​​feature maps into the MaxPool and AvgPool branches respectively to obtain two feature maps, and then perform a Concat splicing operation on them, and then perform convolution and sigmoid to obtain the weight after the fusion of the deep and shallow features, and then perform pixel-by-pixel multiplication with the shallow features to obtain the feature map and perform pixel-by-pixel addition on the shallow feature image to finally obtain the enhanced feature map.

[0029] As a possible implementation, further, the uncertainty modeling in step 3.6) includes the following steps:

[0030] S1: Obtain evidence feature E = [e1, ..., e K ], and ensure that the eigenvalue is greater than 0 by applying the Softplus activation function, the formula is:

[0031] E=SOftplus(Seg)

[0032] S2: Parameterize E as a Dirichlet distribution, the formula is:

[0033]

[0034] Among them, α k and e k are the parameters and evidence of the k-th Dirichlet distribution, (h,w)∈(H,W) and H and W are the width and height of the coarse segmentation graph Seg respectively;

[0035] S3: Calculate the confidence quality and the corresponding uncertainty score, the formula is:

[0036]

[0037] in u (h,w) denote the probability and overall uncertainty of the corresponding class at each pixel coordinate, respectively; represents the Dirichlet intensity, K represents the total number of pixel classes;

[0038] F={u (h,w) ,(h,w)∈(H,W)} is expressed as the uncertainty of each pixel in the segmentation feature map.

[0039] As a possible implementation, further, in step 3.6), a reliable mask M is generated by using the coarse segmentation image Seg, the evidence feature map E, and the uncertainty feature map F, as follows:

[0040] After adding the coarse segmentation map Seg and the evidence feature map E, they are mapped to e with the uncertainty feature map F -F Feature map F on the function w The multiplication results in a reliable mask M.

[0041] As a possible implementation, further, step 4) specifically includes the following steps:

[0042] The preprocessed and amplified image is fed into the neural network constructed in step 3) for network training, and the loss between the network segmentation result and the manually annotated segmentation result is used to guide the network training. When the loss is reduced to meet the requirements or the preset number of iterations is reached, the network training is stopped to obtain the fundus vascular segmentation network model;

[0043] The fundus image to be segmented is directly sent into the trained fundus vascular segmentation network model to obtain the final fundus vascular segmentation result.

[0044] The present invention constructs an attention network based on uncertainty and shallow feature enhancement. Based on the fundus vascular segmentation network, it reduces the downsampling step by one and performs three downsampling steps in total to reduce the depth of the model and improve the segmentation efficiency of the model.

[0045] Secondly, an attention module and a shallow feature enhancement module are added to the network to enhance the expression of shallow features, and an uncertainty modeling module is used before the final output of the network to reduce the adverse effects of uncertainty.

[0046] In the encoding part, the attention module and the shallow feature enhancement module are added to the feature maps of the first three layers before they are spliced ​​with the upsampled feature maps; in the decoding part, the second and third layer feature maps obtained after upsampling are upsampled and convolved by 2 times and 4 times respectively, and then spliced ​​with the first layer feature map and input into the attention module and uncertainty modeling module; the attention module enables the network to focus on the vascular area of ​​the fundus image and improve the network's segmentation accuracy. The shallow feature enhancement module can use the deep feature map to enable the shallow feature map to obtain deep features without losing the details of the shallow features; the uncertainty modeling method can model the uncertainty of the image and combine the deep learning method to reduce high uncertainty to obtain a more refined segmented image.

[0047] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0048] The present invention provides a novel retinal fundus image segmentation method for automatically segmenting retinal blood vessels from fundus images. The method reduces the uncertainty of the model's final output by enhancing shallow feature expression through an attention mechanism and deep feature enhancement, and combines uncertainty modeling with a deep learning approach, thereby improving the accuracy of retinal blood vessel segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 It is a simplified flow chart of the method of the present invention. DETAILED DESCRIPTION

[0051] The present invention will be described in further detail below with reference to the accompanying drawings and examples. It is particularly noted that the following examples are intended only to illustrate the present invention and are not intended to limit the scope of the present invention. Similarly, the following examples are only some embodiments of the present invention and are not intended to be exhaustive. All other embodiments obtained by those of ordinary skill in the art without creative effort are intended to fall within the scope of protection of the present invention.

[0052] Refer to the attached Figure 1 As shown, this embodiment provides a fundus image segmentation method based on uncertainty and shallow feature enhanced attention network, which specifically includes the following steps:

[0053] Step 1): Select fundus retinal image data as training and test sets. Each fundus retinal image sample includes the original color fundus retinal image and the corresponding blood vessel segmentation label. In this example, three public datasets, DRIVE, CHASEDB1, and STARE, are used. DRIVE uses 20 images as the training set and 20 images as the test set; CHASEDB1 uses 21 images as the training set and 7 images as the test set; STARE uses 15 images as the training set and 5 images as the test set.

[0054] Step 2): Preprocess the fundus retinal images in the training set for input into the neural network model. In this embodiment, the preprocessing step includes: grayscaling and normalizing each fundus retinal image, then adjusting the overall contrast using contrast-limited adaptive histogram equalization (CLAHE) and gamma correction with a gamma value of 1.5; and dividing each fundus retinal image and the corresponding retinal vascular segmentation label into smaller blocks or sub-images. Specifically, the training set image data is randomly cropped into 100,000 images with a width and height of 48*48.

[0055] Step 3): Build a neural network model on Pytorch, which includes the following steps:

[0056] Step 3.1): Input the original image into the network and obtain the features through a convolution + BN + ReLU operation Figure 1 , and then undergo three consecutive downsampling and convolution + BN + ReLU operations to obtain feature maps 2, 3, and 4, a total of four feature maps;

[0057] Step 3.2): The features Figure 1 ~3 inputs the attention module to obtain attention feature maps a1, a2, a3;

[0058] Step 3.3): Input the attention feature map a3 and feature map 4 into the shallow feature enhancement module, and output feature map A3; input feature map A3 and attention feature map a2 into the shallow feature enhancement module, and output feature map A2; input feature map A2 and attention feature map a1 into the shallow feature enhancement module, and output feature map A1;

[0059] Step 3.4): After upsampling feature map 4, it is spliced ​​with feature map A3 on the channel, and then subjected to convolution + BN + ReLU operations to obtain feature map U3; after upsampling feature map U3, it is spliced ​​with feature map A2 on the channel, and then subjected to convolution + BN (Batch Normalization) + ReLU operations to obtain feature map U2; after upsampling feature map U2, it is spliced ​​with feature map A1 on the channel, and then subjected to convolution, BN, and ReLU operations to obtain feature map U1;

[0060] Step 3.5): Perform upsampling and convolution operations on feature maps U2 and U3 by 2 and 4 times respectively to make their shapes the same as feature map U1. Then, concatenate feature maps U1, U2, and U3 on the channel to obtain feature map C. Input feature map C into the attention module to obtain feature map C1. Finally, perform residual addition operation on feature map C and feature map C1 to obtain the coarse segmentation map Seg.

[0061] Step 3.6): Perform uncertainty modeling on the coarse segmentation map Seg to obtain the evidence feature map E and the uncertainty feature map F. Generate a reliable mask M by combining the coarse segmentation image Seg with the evidence feature map E and the uncertainty feature map F. Finally, concatenate M and the coarse segmentation map Seg on the channel and perform a 1×1 convolution to output the final refined segmentation result.

[0062] In this embodiment, the attention module consists of a channel attention module and a spatial attention module;

[0063] The channel attention module enables the network to encode the importance of channels, obtain weights for different channels, and thus focus network learning on important channels to improve network learning performance. It is used to pass the input feature map through two parallel MaxPool and AvgPool layers, reducing the feature map from C×H×W to C×1×1. It then passes through the Share MLP module, where it first compresses the number of channels to 1 / r (reduction rate) times the original number of channels, and then expands it to the original number of channels. The module then passes through the ReLU activation function to obtain two activated results. These two output results are added element-by-element, then passed through a sigmoid activation function to obtain the output of Channel Attention, which is then multiplied by the original image to return to the size of C×H×W.

[0064] The spatial attention module enables the network to encode the importance of spatial regions on the feature map, obtain weight values ​​for different regions, and then focus the network's learning on important spatial regions to improve the performance of network learning. It is used to obtain two 1×H×W feature maps through maximum pooling and average pooling of the output results of the channel attention module, and then concatenate the two feature maps through the Concat operation, and then obtain four feature maps through four dilated convolutions with different dilation rates. The four feature maps are then concatenated through the Concat operation, and then a sigmoid operation is performed to obtain the feature map of spatial attention. Finally, the output result is multiplied by the original image to return to the size of C×H×W.

[0065] The shallow feature enhancement module can promote the fusion of deep features and guide the enhanced representation of shallow features. It is used to concatenate shallow features and deep features after convolution and BN, and then input the concatenated feature maps into the MaxPool and AvgPool branches to obtain two feature maps for concat operation. After convolution and sigmoid operation, the weights of deep and shallow feature fusion are obtained. The feature map obtained by pixel-by-pixel multiplication with the shallow feature is added pixel-by-pixel to the shallow feature image to finally obtain the enhanced feature map.

[0066] In this embodiment, the uncertainty modeling in step 3.6) includes the following steps:

[0067] S1: Obtain evidence feature E = [e1, ..., e K ], and ensure that the eigenvalue is greater than 0 by applying the Softplus activation function, the formula is:

[0068] E=Softplus(Seg)

[0069] S2: Parameterize E as a Dirichlet distribution, the formula is:

[0070]

[0071] Among them, α k and e k are the parameters and evidence of the k-th Dirichlet distribution, (h,w)∈(H,W) and H and W are the width and height of the coarse segmentation graph Seg respectively;

[0072] S3: Calculate the confidence quality and the corresponding uncertainty score, the formula is:

[0073]

[0074] in u (h,w)denote the probability and overall uncertainty of the corresponding class at each pixel coordinate, respectively; represents the Dirichlet intensity, K represents the total number of pixel classes;

[0075] F={u (h,w) ,(h,w)∈(H,W)} is expressed as the uncertainty of each pixel in the segmentation feature map.

[0076] In this embodiment, in step 3.6), a reliable mask M is generated by using the coarse segmentation image Seg, the evidence feature map E, and the uncertainty feature map F, as follows:

[0077] After adding the coarse segmentation map Seg and the evidence feature map E, they are mapped to e with the uncertainty feature map F -F Feature map F on the function w The multiplication results in a reliable mask M.

[0078] Step 4): Use the trained neural network model to segment the test set and obtain the final segmentation result; specifically, the following steps are included:

[0079] The preprocessed and amplified image is fed into the neural network constructed in step 3) for network training. The loss between the network segmentation result and the manually annotated segmentation result (vessel segmentation ground truth map) is used to guide network training. When the loss is reduced to meet the requirements or the preset number of iterations is reached, the network training is stopped to obtain the fundus vessel segmentation network model.

[0080] The fundus image to be segmented is directly sent into the trained fundus vascular segmentation network model to obtain the final fundus vascular segmentation result.

[0081] The above descriptions are only some embodiments of the present invention and do not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made by using the contents of the description and drawings of the present invention, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A fundus image segmentation method based on uncertainty and shallow feature enhanced attention network, characterized in that: The steps include: Step 1): Select fundus retinal image data as training set and test set, each fundus retinal image sample includes the original color fundus retinal image and the corresponding blood vessel segmentation label; Step 2): Preprocess the fundus retinal images in the training set; Step 3): Build a neural network model on Pytorch; specifically include the following steps: Step 3.1): Input the original image into the network and perform one convolution, BN, and ReLU operation to obtain feature map 1. Then, perform three consecutive downsampling and convolution, BN, and ReLU operations to obtain feature maps 2, 3, and 4, a total of four feature maps; Step 3.2): Input feature maps 1 to 3 into the attention module to obtain attention feature maps a1, a2, and a3; Step 3.3): Input the attention feature map a3 and feature map 4 into the shallow feature enhancement module, and output feature map A3; input feature map A3 and attention feature map a2 into the shallow feature enhancement module, and output feature map A2; input feature map A2 and attention feature map a1 into the shallow feature enhancement module, and output feature map A1; Step 3.4): After upsampling feature map 4, it is spliced ​​with feature map A3 on the channel, and then subjected to convolution, BN, and ReLU operations to obtain feature map U3; after upsampling feature map U3, it is spliced ​​with feature map A2 on the channel, and then subjected to convolution, BN, and ReLU operations to obtain feature map U2; after upsampling feature map U2, it is spliced ​​with feature map A1 on the channel, and then subjected to convolution, BN, and ReLU operations to obtain feature map U1; Step 3.5): Perform upsampling and convolution operations on feature maps U2 and U3 by 2 and 4 times respectively to make their shapes the same as feature map U1. Then, concatenate feature maps U1, U2, and U3 on the channel to obtain feature map C. Input feature map C into the attention module to obtain feature map C1. Finally, perform residual addition operation on feature map C and feature map C1 to obtain the coarse segmentation map Seg. Step 3.6): Perform uncertainty modeling on the coarse segmentation map Seg to obtain the evidence feature map E and the uncertainty feature map F. Generate a reliable mask M by combining the coarse segmentation image Seg with the evidence feature map E and the uncertainty feature map F. Finally, M and the coarse segmentation map Seg are spliced ​​on the channel and then pass through 1 After convolution of 1, the final refined segmentation result is output; Uncertainty modeling includes the following steps: S1: Get evidence feature E = [ ,..., ], and ensure that the eigenvalue is greater than 0 by applying the Softplus activation function, the formula is: S2: Parameterize E as a Dirichlet distribution, the formula is: in, and are the k-th Dirichlet distribution parameter and evidence, And H and W are the width and height of the coarse segmentation graph Seg respectively; S3: Calculate the confidence quality and the corresponding uncertainty score, the formula is: ， in 、 denote the probability and overall uncertainty of the corresponding class at each pixel coordinate, respectively; represents the Dirichlet intensity, K represents the total number of pixel classes; Expressed as the uncertainty of each pixel in the segmentation feature map; Step 4): Use the trained neural network model to segment the test set and obtain the final segmentation result.

2. The fundus image segmentation method based on uncertainty and shallow feature enhanced attention network according to claim 1 is characterized in that Step 2) specifically includes the following steps: Each fundus retinal image was grayscaled and normalized, and then gamma correction was used with CLAHE and a gamma value of 1.5 to adjust the overall contrast; Each fundus retinal image and the corresponding retinal vessel segmentation labels are divided into smaller blocks or sub-images.

3. The fundus image segmentation method based on uncertainty and shallow feature enhanced attention network according to claim 2 is characterized in that Each fundus retinal image and the corresponding retinal blood vessel segmentation label are divided into smaller blocks or sub-images, specifically: The training set image data is randomly cropped into 100,000 images with a width and height of 48*48.

4. The fundus image segmentation method based on uncertainty and shallow feature enhanced attention network according to claim 1 is characterized in that The attention module consists of a channel attention module and a spatial attention module; The channel attention module is used to pass the input feature map through two parallel MaxPool layers and AvgPool layers, and convert the feature map from C H W becomes C 1 1, and then processed by the Share MLP module; the Share MLP module is used to compress the number of channels to 1 / r times the original number, and then expand it to the original number of channels, and obtain two activated results through the ReLU activation function; Next, add the two output results element by element, and then use a sigmoid activation function to get the output of ChannelAttention, and then multiply this output by the original image to return to C H The size of W; The spatial attention module is used to perform maximum pooling and average pooling operations on the output results of the channel attention module to obtain two 1 H The feature map of W is then concatenated by the Concat operation, and four feature maps are obtained by four dilated convolutions with different dilation rates. The four feature maps are then concatenated by the Concat operation, and a sigmoid operation is performed to obtain the feature map of spatial attention. Finally, the output result is multiplied by the original image to return to C. H W size.

5. The fundus image segmentation method based on uncertainty and shallow feature enhanced attention network according to claim 1, characterized in that The shallow feature enhancement module is used to perform a Concat operation on the shallow features and deep features after convolution and BN, and then input the spliced ​​feature maps into the MaxPool and AvgPool branches respectively to obtain two feature maps, which are then subjected to a Concat operation and then subjected to convolution and sigmoid to obtain the weight after the fusion of deep and shallow features. The feature map obtained by pixel-by-pixel multiplication with the shallow features is then added pixel-by-pixel to the shallow feature image to finally obtain the enhanced feature map.

6. The fundus image segmentation method based on uncertainty and shallow feature enhanced attention network according to claim 1, characterized in that: In step 3.6), a reliable mask M is generated by coarsely segmenting the image Seg, the evidence feature map E, and the uncertainty feature map F as follows: After adding the coarse segmentation map Seg and the evidence feature map E, they are mapped to the uncertainty feature map F Feature graph on function The multiplication results in a reliable mask M.

7. The fundus image segmentation method based on uncertainty and shallow feature enhanced attention network according to claim 1 is characterized in that Step 4) specifically includes the following steps: The preprocessed and amplified image is fed into the neural network constructed in step 3) for network training. The loss between the network segmentation results and the manually annotated segmentation results is used to guide network training. When the loss is reduced to meet the requirements or the preset number of iterations is reached, the network training is stopped to obtain the fundus vessel segmentation network model. The fundus image to be segmented is directly sent into the trained fundus vascular segmentation network model to obtain the final fundus vascular segmentation result.

Images