Retinal image segmentation method based on residual attention and multi-scale feature aggregation

Abstract

The present invention relates to the field of image processing technology, and in particular to a retinal image segmentation method based on residual attention and multi-scale feature aggregation, comprising: taking a U-Net network structure as a basic framework, introducing a plurality of improved residual modules in an encoder part and a decoder part of a network respectively, and connecting a channel attention module after each of the improved residual modules to reduce the problem of information loss in the convolution and pooling processes, strengthen important information, and suppress redundant information; introducing a multi-scale feature aggregation module in a skip path to effectively fuse and extract vascular features of different scales, thereby improving the feature extraction capability of the network and providing richer information for a decoding layer to reconstruct a fundus image.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and more specifically to a retinal image segmentation method based on residual attention and multi-scale feature aggregation. Background Art

[0002] The vascular structure and morphology of fundus images contain rich information. Retinal image segmentation plays a vital role in assisting the screening of eye diseases. However, due to the low contrast between the tiny blood vessels and the background in fundus images and the complex tree-like branching structure of blood vessels, accurate retinal image segmentation is still a challenging task.

[0003] In view of the complexity of the retinal vascular topology and the problems of low contrast and uneven light that may occur during fundus data collection, traditional machine learning segmentation algorithms often show low segmentation accuracy and weak anti-interference ability. The convolutional neural network algorithm based on deep learning has significantly improved the performance of segmentation tasks due to its high accuracy, powerful feature extraction ability and robustness. However, the traditional U-Net model is relatively simple in feature extraction and is difficult to adapt to the variability of vascular scale. In addition, its jump connection simply fuses information in the encoding and decoding stage, which may lead to redundancy in the learned features, thus affecting the segmentation accuracy. Summary of the invention

[0004] In view of this, the present invention provides a retinal image segmentation method based on residual attention and multi-scale feature aggregation, which can achieve more accurate retinal vascular image reconstruction.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides a retinal image segmentation method based on residual attention and multi-scale feature aggregation, comprising:

[0007] Taking the U-Net network structure as the basic framework, multiple improved residual modules are introduced into the encoder part and the decoder part of the network respectively, a channel attention module is connected after each of the improved residual modules to form a channel attention residual feature extraction module, and a multi-scale feature aggregation module is introduced in the skip path to obtain a retinal image segmentation model;

[0008] In the encoder part of the retinal image segmentation model, the improved residual module uses a short connection mechanism to extract vascular feature information of different scales from the downsampled retinal feature map; the channel attention module learns the importance weights of the feature channels output by the residual module and dynamically reshapes the feature channels to strengthen key information and suppress redundant information;

[0009] Aggregating the extracted blood vessel feature information of different scales based on the multi-scale feature aggregation module;

[0010] In the decoder part of the retinal image segmentation model, the retinal feature map processed by the multi-scale feature aggregation module is fused with the retinal feature map upsampled in the decoding stage to reconstruct the retinal vascular image.

[0011] Furthermore, the improved residual module processes the data stream by:

[0012] The data flow starts with a 3×3 convolution, and is processed by batch normalization, Dropout, and ReLU activation functions in sequence. It then enters another 3×3 convolution, and is processed by BN and Dropout again. After that, the input feature map is fused with the output feature map of the current layer through identity mapping. Finally, the fused feature map is activated by the ReLU activation function to generate the final output of the residual module.

[0013] Furthermore, the step of reshaping the feature channel based on the channel attention module includes:

[0014] The feature map X output by the residual module is pooled by a global average pooling operation r-o Compressed into a channel descriptor with a global receptive field, the spatial information of each channel is compressed into a single value, and a feature description of size 1×1×C is obtained;

[0015] Perform an excitation operation on the compressed features, and learn the importance weight X of each channel through one or more layers of fully connected networks. se-e ;

[0016] The learned channel importance weight X se-e With the original feature map X r-o Perform element-by-element multiplication to obtain the weighted feature map.

[0017] Furthermore, the encoder part of the retinal image segmentation model includes a five-layer channel attention residual feature extraction module, which performs feature extraction at different scales to obtain four inputs input1, input2, input3 and input4 of the multi-scale feature aggregation module and an input input5 of the decoder.

[0018] Furthermore, the multi-scale feature aggregation module processes the multi-scale features as follows:

[0019] Input1 passes through a 3×3 convolution and is downsampled to obtain Input1'; Input2 passes through a 3×3 convolution and is downsampled to obtain Input2'; Input3 passes through a 3×3 convolution and is processed through a channel attention residual feature extraction module to obtain Input3'; Input4 passes through a 3×3 convolution and is upsampled to obtain Input4';

[0020] Connect Input1', Input2', Input3' and Input4' to get feature F Fusion ;

[0021] For feature F Fusion 1×1 convolution, 3×3 convolution with Rate=4, 3×3 convolution with Rate=8, 3×3 convolution with Rate=12, and pooling operations are performed respectively, and upsampling, channel attention residual feature extraction module, and downsampling are performed to obtain output1, output2, output3, output4, and output5.

[0022] Furthermore, in the decoder part of the retinal image segmentation model, the process of performing feature fusion includes:

[0023] After being processed by the channel attention residual feature extraction module, iutput5 is merged with output5 to obtain output5', output5' is concatenated with output4, and then processed by the channel attention residual feature extraction module to obtain output4';

[0024] Concatenate output4' and output3 and process them with the channel attention residual feature extraction module to get output3';

[0025] Concatenate output3' and output2 and process them with the channel attention residual feature extraction module to get output2';

[0026] Concatenate output2' and output1 and process them with the channel attention residual feature extraction module to get output1';

[0027] Perform 1×1 convolution on output1' to get the final output.

[0028] Furthermore, the loss function expression of the retinal image segmentation model is:

[0029] L = ∝Loss BCE +(1-∝)Loss Dice

[0030] Among them, Loss BCE Represents binary cross loss, Loss Dice represents the Dice loss, and ∝ represents the weight of the binary cross loss.

[0031] In a second aspect, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, the steps of the above-mentioned retinal image segmentation method based on residual attention and multi-scale feature aggregation are implemented.

[0032] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the above-mentioned retinal image segmentation method based on residual attention and multi-scale feature aggregation are implemented.

[0033] It can be seen from the above technical solutions that, compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention adds a multi-scale feature aggregation module on the basis of U-Net, which can capture contextual information at different levels and effectively integrate vascular features of different scales, showing higher accuracy and robustness. In addition, the introduction of attention mechanism and improved residual module can alleviate the information loss problem caused by pooling and convolution in the feature extraction process, retain more important information, and suppress redundant information. The features after multi-scale feature aggregation processing are fused with the features of the decoding layer to achieve accurate restoration of vascular details. Compared with the existing methods, this method significantly improves the sensitivity and accuracy of retinal image segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0036] Figure 1 A schematic diagram of the structure of the retinal image segmentation model provided by the present invention;

[0037] Figure 2 A schematic diagram of the structure of an improved residual module provided by the present invention;

[0038] Figure 3 A schematic diagram of the structure of the multi-scale feature aggregation module provided by the present invention. DETAILED DESCRIPTION

[0039] 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.

[0040] like Figure 1 As shown, the embodiment of the present invention discloses a retinal image segmentation method based on residual attention and multi-scale feature aggregation, including:

[0041] Taking the U-Net network structure as the basic framework, multiple improved residual modules are introduced into the encoder and decoder parts of the network respectively. A channel attention module is connected after each improved residual module to form a channel attention residual feature extraction module. A multi-scale feature aggregation module is introduced in the skip path to obtain a retinal image segmentation model.

[0042] In the encoder part of the retinal image segmentation model, each improved residual module uses a short connection mechanism to extract vascular feature information of different scales from the downsampled retinal feature map; the channel attention module dynamically reshapes the feature channel by learning the weights of the feature channel to strengthen the key information and suppress the redundant information;

[0043] The extracted vascular feature information of different scales is aggregated based on the multi-scale feature aggregation module;

[0044] In the decoder part of the retinal image segmentation model, the retinal feature map processed by the multi-scale feature aggregation module is fused with the retinal feature map upsampled in the decoding stage to reconstruct the retinal vascular image.

[0045] The retinal image segmentation model (MFAM-Net) constructed by the present invention performs accurate segmentation of fundus retinal images in an end-to-end, pixel-by-pixel manner. The overall architecture of the model aims to effectively improve the performance of segmentation tasks by combining the channel attention mechanism, residual module, and multi-scale feature aggregation module.

[0046] The basic framework of the model is a variant of the U-Net network structure, which is divided into two parts: an encoder and a decoder. The encoder part includes five channel attention residual feature extraction modules, each of which is a feature extraction layer for deepening feature extraction and enhancing the learning ability of the model, and four downsampling layers for gradually reducing the spatial dimension of the image while increasing the depth of the feature map. In the encoder, compared with the encoding block of the U-Net network structure, the present invention introduces a short connection mechanism of the residual encoding block, which not only effectively prevents the gradient vanishing problem and accelerates the convergence speed of the network, but also strengthens the feature transfer ability by integrating the channel attention module, and promotes the network to learn important feature information. In addition, after the first four feature extraction layers of the network, the shallow and deep information of the context is aggregated by adopting a multi-scale feature aggregation module in the jump path, thereby capturing semantic information of different scales and realizing the comprehensive capture of vascular pixels. The decoder part of the present invention combines the feature map after multi-scale feature aggregation processing in the jump path with the feature map upsampled in the decoding stage to achieve effective fusion of deep and shallow features and better reconstruct image detail information. The present invention uses Residual Block modules in both the encoder and the decoder to extract richer retinal image features, and processes the output of each residual module through a channel attention (Squeeze-and-Excitation, SE) module, which can selectively emphasize important feature information while suppressing redundant information.

[0047] The present invention improves the design of the U-Net network structure, combines a multi-scale feature aggregation module, an improved residual module and a channel attention mechanism, so that the final model can effectively improve the accuracy and overall segmentation performance of retinal image segmentation.

[0048] The above modules are further explained below.

[0049] (1) Channel attention residual feature extraction module

[0050] The model uses multiple channel attention residual feature extraction modules, and the structure of each channel attention residual feature extraction module is a residual module followed by a channel attention module.

[0051] 1) Residual Module

[0052] The design concept of the residual network (ResNet) originated from solving the degradation problem of deep neural networks after increasing the number of layers, that is, increasing the network depth leads to worse training results. This phenomenon shows that although deep networks have stronger representation capabilities in theory, they are difficult to optimize during actual training. The core idea of ​​the residual network is to directly transfer low-level feature information to high-level networks by introducing identity mapping, ensuring that the deep network can at least reach the performance level of the shallow network, while alleviating the gradient vanishing and gradient exploding problems caused by increasing network depth. Through such a structural design, the residual network greatly improves the training efficiency and performance of deep networks, becoming an important milestone in the field of deep learning and is widely used in a variety of tasks such as image segmentation and object detection.

[0053] In the residual network, each residual block is designed with two paths: one is the path of the residual function F(X), and the other is the path of the identity mapping X, which is also called a short cut. The main difference between this structure and the traditional convolutional layer is that it adds an identity mapping connection outside the main path. This design makes the input X r-i Some layers can be skipped directly through short connections to pass to deeper layers, thereby increasing the depth of the network without adding additional parameters. As shown in the following formula:

[0054] H(X r-i )=F(X r-i )+X r-i

[0055] Among them, X r-i is the input of the improved residual module, F(X r-i ) is the feature variation learned by the residual function, H(X r-i ) is the output feature map, denoted by X r-o Here F(X r-i )+X r-i It is an element-wise addition operation, requiring F(X r-i ) and X r-i The dimensions must match.

[0056] like Figure 2As shown, the processing flow of each residual module on the data stream is: starting from a 3×3 convolution, followed by a series of processing of batch normalization (BN), Dropout and ReLU activation function, and then the data stream enters another 3×3 convolution, and is processed by BN and Dropout again. Subsequently, the input feature map is fused with the output feature map of the current layer through identity mapping. Finally, the fused feature map is activated by the ReLU activation function to generate the final output of the residual module, and the output of the residual module is used as the input of the channel attention module. In the network encoding and decoding layer of the present invention, the traditional convolution layer is replaced by an improved residual module, which effectively alleviates the occurrence of overfitting and improves the generalization ability of the model.

[0057] 2) Channel Attention Module

[0058] The visual attention mechanism is an efficient way for the human brain to process visual information. It allows the brain to quickly focus on key information in a complex environment while ignoring irrelevant background information. Inspired by the human visual attention mechanism, the field of deep learning has also successfully simulated this phenomenon through technology. In deep learning tasks, the attention mechanism prioritizes computing resources to more important features, effectively improving the performance of the model in recognition and analysis tasks. This mechanism makes the model more accurate and efficient when processing complex data by enhancing the network's ability to respond to key information.

[0059] For example, in face recognition tasks, the network needs to pay special attention to key facial features (such as ears, eyes, mouth, etc.), while paying relatively less attention to background information. This mechanism allows the network to effectively identify and process key visual information with limited resources, improving processing efficiency and accuracy.

[0060] The channel attention mechanism is a type of attention mechanism that strengthens important features and suppresses minor features by evaluating the contribution of different channels (or feature maps) to the final output. This is achieved through automatic learning of the network, which learns features that are more important than other features and redistributes the weight coefficients of the channels accordingly.

[0061] Specifically, the channel attention module of the present invention uses three operations of compression, excitation and reweighting to recalibrate the weights of the feature channels, and the output feature map X of the residual module is r-o The following processing is performed, the specific steps are:

[0062] Compression: The feature map output by the residual module is compressed into a channel descriptor with a global receptive field through a global average pooling operation, and the spatial information of each channel is compressed into a single value to obtain a feature description of size 1×1×C; where the feature map X r-oThe dimension is H×W×C, where H, W, and C represent the height, width, and number of channels, respectively. The function of the compression part is to compress the dimension from H×W×C to 1×1×C. The compressed feature is recorded as X se-s .

[0063]

[0064] Among them, F sq (·) indicates a compression operation.

[0065] Stimulation: The compressed feature X se-s Perform the excitation operation, including two fully connected layers and two activation functions σ and δ, and learn the importance weight X of each channel se-e :

[0066] X se-e =F ex (X se-s ,W)=σ(g(X se-s ,W))=σ(W 2 δ(X se-s ,W))

[0067] Reweighting: Reweight the learned channel importance weights X se-e With the original feature map X r-o Perform element-by-element multiplication to obtain the weighted feature map X se :

[0068] X se =F scale (X se-e ,X r-o )=X se-e ·X r-o

[0069] In order to effectively improve the feature extraction capability of retinal images and improve the efficiency of the network in feature utilization, the present invention adopts a strategy of combining residual modules and channel attention modules. By connecting a channel attention module after each improved residual module, the feature channels are further dynamically reshaped by learning the weights of the feature channels. By learning the importance of each channel, the ability to capture key information is effectively enhanced while reducing the interference of unimportant features. This strategy of combining residual modules and channel attention modules can not only improve the network's sensitivity to detail features, but also improve the accuracy and efficiency of the overall model in extracting retinal image features.

[0070] (2) Multi-scale feature processing

[0071] In order to obtain a good segmentation effect, the present invention extracts multi-scale features in the form of a channel attention residual module, and the process is as follows:

[0072] The first layer of feature extraction: The input is a 96×96 pixel image block I, which is extracted into a feature F1 of size 96×96×32 through the channel attention residual module, i.e., input 1 of the multi-scale feature aggregation module;

[0073] The second layer feature extraction: The input is F1, which is extracted by the maximum pooling and channel attention residual feature extraction modules to obtain the feature F2 of size 48×48×64, that is, the input 2 of the multi-scale feature aggregation module, input2;

[0074] The third layer feature extraction: The input is F2, which is extracted by the maximum pooling and channel attention residual feature extraction modules to obtain the feature F3 of size 24×24×128, that is, the input 3 of the multi-scale feature aggregation module, input3;

[0075] The fourth layer feature extraction: The input is F3, which is extracted by the maximum pooling and channel attention residual feature extraction modules to obtain the feature F4 of size 12×12×256, that is, the input 4 of the multi-scale feature aggregation module;

[0076] Fifth layer feature extraction: The input is F4, which is extracted by the maximum pooling and channel attention residual feature extraction modules to obtain feature F5 of size 6×6×512, input 5 of the encoding stage.

[0077] (3) Multi-scale feature aggregation module

[0078] In view of the complexity, morphological diversity and uneven thickness of retinal vascular structure, as well as the fact that the noise of image acquisition and lesions may interfere with the accuracy of vascular segmentation, it is necessary to correctly capture and utilize contextual information to improve segmentation accuracy. Especially when processing fundus images, if bleeding, spots and infiltration caused by lesions are misclassified as vascular pixels due to insufficient contextual information, the accuracy of the segmentation results will be seriously affected. To solve this problem, the present invention proposes a multi-scale feature aggregation module based on atrous spatial pyramid pooling to capture semantic information at different scales. Figure 3 As shown. For multi-scale features input1, input2, input3 and input4, the processing flow of the multi-scale feature aggregation module is as follows:

[0079] Input1 passes through a 3×3 convolution and is downsampled to obtain Input1'; Input2 passes through a 3×3 convolution and is downsampled to obtain Input2'; Input3 passes through a 3×3 convolution and is processed through the channel attention residual feature extraction module to obtain Input3'; Input4 passes through a 3×3 convolution and is upsampled to obtain Input4'.

[0080] Connect Input1', Input2', Input3', Input4' to get feature F Fusion , F Fusion The multi-scale features in the skip path are effectively aggregated. The skip path here refers to taking the output of one layer as the input of the next layer and the number of layers through which the extracted features pass is different.

[0081] For feature F Fusion 1×1 convolution, 3×3 convolution with Rate=4, 3×3 convolution with Rate=8, 3×3 convolution with Rate=12, and pooling operations are performed respectively, and upsampling, channel attention residual feature extraction module, and downsampling are performed to obtain output1, output2, output3, output4, and output5.

[0082] The goal of MFAM is to enhance the network's ability to capture vascular context information by fusing feature information at different scales, and to establish long-distance dependencies between blood vessels, allowing the network to not only identify detailed features in the image, but also understand the overall structure and surrounding environment of the blood vessels at a more macro level, thereby significantly improving the performance of blood vessel segmentation. By aggregating features at different levels, MFAM enables the network to adapt to various scales and forms of blood vessels, and effectively distinguish between blood vessels and non-vascular structures, especially in complex or ambiguous areas. In this way, MFAM not only improves the accuracy of segmentation, but also enhances the model's robustness to vascular features in complex backgrounds.

[0083] (4)Decoder Processing

[0084] The processing of the decoder is mainly the fusion of deep and shallow features, specifically:

[0085] iutput5 is processed by the channel attention residual module and merged with output5 to obtain output5', output5' is concatenated with output4 and processed by the channel attention residual feature module to obtain output4';

[0086] Concatenate output4' and output3 and process them with the channel attention residual feature module to get output3';

[0087] Concatenate output3' and output2 and process them with the channel attention residual feature extraction module to get output2';

[0088] Concatenate output2' and output1 and process them with the channel attention residual feature extraction module to get output1';

[0089] Perform 1×1 convolution on output1' to obtain the final output result graph.

[0090] (5) Loss function design

[0091] The loss function expression of the retinal image segmentation model is:

[0092] L = ∝Loss BCE +(1-∝)Loss Dice

[0093] Among them, Loss BCE Represents binary cross loss, Loss Dice represents the Dice loss, and ∝ represents the weight of the binary cross loss.

[0094] Among them, the Binary Cross Entropy (BCE) loss function is particularly suitable for processing pixel-level classification tasks, such as the segmentation of retinal blood vessels and background. This segmentation task is essentially to classify each pixel of the input image and determine whether it belongs to a blood vessel (i.e., a positive sample) or a non-vascular (i.e., a negative sample) pixel, thereby achieving accurate segmentation of the retinal image. When using the binary cross entropy loss function for fundus image segmentation, the goal of the model prediction is to minimize the difference between the predicted pixel value and the actual label value. The loss function is calculated as follows:

[0095]

[0096] Among them, g i Represents the true label value, which can be 0 or 1. i Represents the predicted value of the pixel. The first part of the formula g i logp i Corresponding to the loss of positive samples, the second part (1-g i )log(1-p i ) corresponds to the loss of negative samples. i When the value is 0, in order to reduce the loss value, the model prediction value p i Need to be close to 0; otherwise, when g i When the value is 1, in order to make the loss value smaller, p i It should be as close to 1 as possible. This mechanism ensures that the model continuously adjusts parameters during training to more accurately predict vascular and non-vascular pixels, thereby improving the accuracy of segmentation.

[0097] The Dice loss function is a key performance indicator widely used in medical image segmentation tasks. It is particularly used to evaluate the similarity between the model's predicted segmentation results and the actual annotations, alleviating the problem of sample imbalance. The Dice coefficient (also known as the F1 score) is a statistical tool for measuring the similarity between two sample sets. The closer its value is to 1, the higher the similarity. The Dice loss function formula is defined as follows:

[0098]

[0099] Among them, X represents the pixel set of the segmentation result predicted by the model, and Y represents the pixel set of the gold standard label manually annotated by professional doctors. X∩Y represents the number of pixels in the intersection of X and Y, and |X| and |Y| are the number of pixels of each.

[0100] The present invention combines these two loss functions to better guide the network training, alleviates the problem of class imbalance, and makes the training process more efficient and stable, taking into account the accuracy and comprehensiveness of the prediction, so as to show better performance in pixel-level image segmentation tasks.

[0101] In other embodiments, the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor; characterized in that when the processor executes the computer program, the steps of the above-mentioned retinal image segmentation method based on residual attention and multi-scale feature aggregation are implemented.

[0102] In some other embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the above-mentioned retinal image segmentation method based on residual attention and multi-scale feature aggregation are implemented.

[0103] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0104] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A retinal image segmentation method based on residual attention and multi-scale feature aggregation, characterized in that: include: Taking the U-Net network structure as the basic framework, multiple improved residual modules are introduced into the encoder part and the decoder part of the network respectively, a channel attention module is connected after each of the improved residual modules to form a channel attention residual feature extraction module, and a multi-scale feature aggregation module is introduced in the skip path to obtain a retinal image segmentation model; In the encoder part of the retinal image segmentation model, the improved residual module uses a short connection mechanism to extract vascular feature information of different scales from the downsampled retinal feature map; The channel attention module dynamically reshapes the feature channels by learning the weights of the feature channels to strengthen key information and suppress redundant information; Aggregating the extracted blood vessel feature information of different scales based on the multi-scale feature aggregation module; In the decoder part of the retinal image segmentation model, the retinal feature map processed by the multi-scale feature aggregation module is fused with the retinal feature map upsampled in the decoding stage to reconstruct a retinal vascular image; The steps of reshaping the feature channel based on the channel attention module include: The feature map X output by the residual module is pooled by a global average pooling operation r-o Compressed into a channel descriptor with a global receptive field, the spatial information of each channel is compressed into a single value, and a feature description of size 1×1×C is obtained; Perform an excitation operation on the compressed features, and learn the importance weight X of each channel through one or more layers of fully connected networks. se-e ; The learned channel importance weight X se-e With the original feature map X r-o Perform element-by-element multiplication to obtain the weighted feature map.

2. The retinal image segmentation method based on residual attention and multi-scale feature aggregation according to claim 1, characterized in that: The improved residual module processes the data stream including: The data flow starts with a 3×3 convolution, and is processed by batch normalization, Dropout, and ReLU activation functions in sequence. It then enters another 3×3 convolution, and is processed by BN and Dropout again. After that, the input feature map is fused with the output feature map of the current layer through identity mapping. Finally, the fused feature map is activated by the ReLU activation function to generate the final output of the residual module.

3. The retinal image segmentation method based on residual attention and multi-scale feature aggregation according to claim 1, characterized in that: The encoder part of the retinal image segmentation model includes a five-layer channel attention residual feature extraction module, which performs feature extraction at different scales to obtain four inputs input1, input2, input3 and input4 of the multi-scale feature aggregation module and an input input5 of the decoder.

4. The retinal image segmentation method based on residual attention and multi-scale feature aggregation according to claim 3, characterized in that: The processing flow of the multi-scale feature aggregation module for multi-scale features is as follows: Input1 passes through a 3×3 convolution and is downsampled to obtain Input1'; Input2 passes through a 3×3 convolution and is downsampled to obtain Input2'; Input3 passes through a 3×3 convolution and is processed through a channel attention residual feature extraction module to obtain Input3'; Input4 passes through a 3×3 convolution and is upsampled to obtain Input4'; Connect Input1', Input2', Input3' and Input4' to get feature F Fusion ; For feature F Fusion 1×1 convolution, 3×3 convolution with Rate=4, 3×3 convolution with Rate=8, 3×3 convolution with Rate=12, and pooling operations are performed respectively, and upsampling, channel attention residual feature extraction module, and downsampling are performed to obtain output1, output2, output3, output4, and output5.

5. The retinal image segmentation method based on residual attention and multi-scale feature aggregation according to claim 4, characterized in that: In the decoder part of the retinal image segmentation model, the process of feature fusion includes: After being processed by the channel attention residual feature extraction module, iutput5 is merged with output5 to obtain output5', output5' is concatenated with output4, and then processed by the channel attention residual feature extraction module to obtain output4'; Concatenate output4' and output3 and process them with the channel attention residual feature extraction module to get output3'; Concatenate output3' and output2 and process them with the channel attention residual feature extraction module to get output2'; Concatenate output2' and output1 and process them with the channel attention residual feature extraction module to get output1'; Perform 1×1 convolution on output1' to get the final output.

6. The retinal image segmentation method based on residual attention and multi-scale feature aggregation according to claim 1, characterized in that: The loss function expression of the retinal image segmentation model is: L=∝Loss BCE +(1-∝)Loss Dice Among them, Loss BCE Represents binary cross loss, Loss Dice represents the Dice loss, and ∝ represents the weight of the binary cross loss.

7. An electronic device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor; characterized in that when the processor executes the computer program, the steps of the retinal image segmentation method based on residual attention and multi-scale feature aggregation as described in any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the retinal image segmentation method based on residual attention and multi-scale feature aggregation as described in any one of claims 1 to 6 are implemented.

Images