A retinal vascular image segmentation method and system based on differential attention

Abstract

The present invention belongs to the field of medical image segmentation, and provides a retinal vascular image segmentation method and system based on differential attention. The method includes obtaining a retinal vascular image; obtaining a retinal fundus vascular image segmentation result based on the retinal vascular image and a differential attention multi-scale residual network; wherein the differential attention multi-scale residual network includes a multi-scale input module, an encoder module, a differential amplification module and a decoder module; the multi-scale input module is used to extract multi-scale information of the retinal vascular image; the encoder module is used to encode the multi-scale information; the differential amplification module is used to extract low-frequency information and high-frequency information of the encoded multi-scale information, and then extract the features of the low-frequency information and the high-frequency information respectively; the attention mechanism is introduced in the decoder module to enhance the attention of the required attention area while suppressing irrelevant areas, and finally restore the extracted low-frequency and high-frequency features to the original resolution.

Description

Technical Field

[0001] The present invention belongs to the field of medical image segmentation, and in particular relates to a retinal vascular image segmentation method and system based on differential attention. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] As a non-invasive and rapid diagnostic method for modern ophthalmic diseases, retinal fundus vessel segmentation is an important part of computer-aided clinical diagnosis of retinal diseases. It can help ophthalmologists diagnose certain retinal diseases, such as retinal vessel occlusion, diabetic retinopathy, hypertension, and glaucoma. The characteristics of these diseases can be obtained by observing the changes in retinal blood vessels, such as shape, width, branching pattern, etc. Early detection and diagnosis can help potential patients with effective prevention and treatment.

[0004] In clinical practice, ophthalmologists usually manually annotate the blood vessels in retinal fundus images, which requires a lot of experience and time to correctly segment these retinal images. However, the size and shape of retinal fundus blood vessels are extremely irregular, which makes the manual segmentation process very cumbersome and error-prone, which is not feasible in clinical applications. The inventors found that most of the current methods to solve the above problems are generally based on algorithms such as image processing, image enhancement or deep learning. Although these methods perform well, they inevitably make the network more complex. When the features accumulate to the limit, it may not be possible to extract the small blood vessel parts, resulting in unsatisfactory segmentation results. Summary of the invention

[0005] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a retinal vascular image segmentation method and system based on differential attention, which segments retinal vascular images based on a differential attention multi-scale residual network, and can achieve high segmentation accuracy while ensuring high robustness, thereby ensuring accurate retinal fundus vascular image segmentation.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A first aspect of the present invention provides a retinal vascular image segmentation method based on differential attention, comprising:

[0008] Acquire retinal vascular images;

[0009] Based on the retinal vascular image and the differential attention multi-scale residual network, the retinal fundus vascular image segmentation result is obtained;

[0010] Among them, the differential attention-based multi-scale residual network includes a multi-scale input module, an encoder module, a differential amplification module and a decoder module; the multi-scale input module is used to extract multi-scale information of retinal vascular images; the encoder module is used to encode the multi-scale information; the differential amplification module is used to extract low-frequency information and high-frequency information of the encoded multi-scale information respectively, and then extract the features of the low-frequency information and the high-frequency information respectively; the attention mechanism is introduced in the decoder module to enhance the attention of the required attention area while suppressing irrelevant areas, and finally restore the extracted low-frequency and high-frequency features to the original resolution.

[0011] A second aspect of the present invention provides a retinal vascular image segmentation system based on differential attention, comprising:

[0012] An image acquisition module, used for acquiring retinal blood vessel images;

[0013] An image segmentation module is used to obtain a retinal fundus vascular image segmentation result based on a retinal vascular image and a differential attention multi-scale residual network;

[0014] Among them, the differential attention-based multi-scale residual network includes a multi-scale input module, an encoder module, a differential amplification module and a decoder module; the multi-scale input module is used to extract multi-scale information of retinal vascular images; the encoder module is used to encode the multi-scale information; the differential amplification module is used to extract low-frequency information and high-frequency information of the encoded multi-scale information respectively, and then extract the features of the low-frequency information and the high-frequency information respectively; the attention mechanism is introduced in the decoder module to enhance the attention of the required attention area while suppressing irrelevant areas, and finally restore the extracted low-frequency and high-frequency features to the original resolution.

[0015] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the retinal vascular image segmentation method based on differential attention as described above.

[0016] A fourth aspect of the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the retinal vascular image segmentation method based on differential attention as described above are implemented.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] The present invention obtains the retinal fundus vascular image segmentation result based on the retinal vascular image and the differential attention multi-scale residual network; wherein the differential attention multi-scale residual network includes a multi-scale input module, an encoder module, a differential amplification module and a decoder module; the multi-scale input module proposed in the present invention takes into account the feature extraction difference caused by the difference in network layer depth. This method retains the original features of the image to a great extent, avoids the feature loss problem that occurs in the up- and down-sampling process of the deep convolutional neural network to a certain extent, and makes the segmentation result more accurate;

[0019] The present invention adopts a differential amplification module to make the network pay more attention to high-frequency and low-frequency information during training, and the boundary areas with high segmentation accuracy are more accurate; the attention module can highlight the significant output features in the encoding path and suppress irrelevant areas in the jump connection. These modules in the differential attention-based multi-scale residual network of the present invention effectively solve the problems of insufficient segmentation accuracy and blurred boundaries of branch vessels and small blood vessels, and at the same time, the background segmentation outside the blood vessels is also more accurate.

[0020] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0022] Figure 1 is a flow chart of a method for segmenting a retinal fundus blood vessel image according to an embodiment of the present invention;

[0023] Figure 2 This is a diagram of a multi-scale residual network architecture based on differential attention according to an embodiment of the present invention;

[0024] Figure 3 This is the maximum pooling operation of the embodiment of the present invention;

[0025] Figure 4(a) shows a common convolution unit;

[0026] Figure 4(b) shows the residual convolution unit;

[0027] Figure 5 is a structural diagram of a differential amplifier module according to an embodiment of the present invention;

[0028] Figure 6 This is a structural diagram of the attention gate module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0030] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0031] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0032] Embodiment 1

[0033] like Figure 1 As shown, this embodiment provides a retinal vascular image segmentation method based on differential attention, which specifically includes the following steps:

[0034] S101: Acquire retinal blood vessel images;

[0035] S102: Based on the retinal vascular image and the differential attention multi-scale residual network, a retinal fundus vascular image segmentation result is obtained;

[0036] Among them, Figure 2 As shown, the differential attention-based multi-scale residual network includes a multi-scale input module, an encoder module, a differential amplification module and a decoder module; the multi-scale input module is used to extract multi-scale information of retinal vascular images; the encoder module is used to encode the multi-scale information; the differential amplification module is used to extract low-frequency information and high-frequency information of the encoded multi-scale information, respectively, and then extract the features of the low-frequency information and the high-frequency information, respectively; the attention mechanism is introduced in the decoder module to enhance the attention of the required attention area while suppressing irrelevant areas, and finally restore the extracted low-frequency and high-frequency features to the original resolution.

[0037] The differential attention multi-scale residual network in this embodiment follows the U-Net architecture, and introduces a differential amplification module and an attention module into the encoder-decoder structure with jump connections. The convolution units in the network are all based on residual connections, which effectively improves the speed of network convergence.

[0038] In this embodiment, the multi-scale input module includes inputs of at least three scales, the input of the first scale is a preprocessed grayscale image of the original training set, and continuous maximum pooling operations are performed on the input of the first scale to obtain inputs of the second scale and inputs of scales after the second scale in turn.

[0039] For example, Figure 2 In the figure, the multi-scale input module contains inputs of four different scales. Input (0) is the preprocessed grayscale image of the original training set, and continuous maximum pooling operations are performed on it to obtain input (1), input (2) and input (3). The purpose of the multi-scale input module is to enable the network to fuse multi-scale information and strengthen the training of the encoder network layer to solve the problem of insufficient segmentation of small blood vessels in retinal fundus images.

[0040] The number of layers included in the encoder module is the same as the number of input scales of the multi-scale input module. Each layer of the encoder module includes a residual convolution unit and a downsampling operation. The residual convolution unit is used to fuse the multi-scale input image to supplement the feature information and compress the image to reduce the image resolution through the downsampling operation.

[0041] For example, Figure 2 In the , the encoder module consists of four layers, each of which includes a residual convolution unit and a downsampling operation. The encoder module fuses multi-scale input images to supplement feature information and compresses the image to reduce the image resolution through downsampling operations.

[0042] like Figure 5 As shown, the differential amplifier module is connected to each layer of the residual convolution unit of the encoder module to form a jump connection. It is used to extract high-frequency and low-frequency information of images with different resolutions. In the differential amplifier module, the low-frequency information is extracted by using the average pooling operation as a low-pass filter, and the difference between the maximum pooling operation and the convolution operation is used as a high-pass filter to extract the high-frequency information. The high-frequency information and the low-frequency information are convolved separately to further extract features.

[0043] Figure 3 is the maximum pooling operation. Taking input (1) as an example, the original input grayscale image input (0) of our dataset is (a). We find the maximum value of all values ​​in each 2×2 non-overlapping area and get (b), which is input (1).

[0044] Figure 4(a) shows a normal convolution unit, and Figure 4(b) shows a residual convolution unit. ⊕ indicates the addition operation of corresponding elements of the matrix.

[0045] The number of layers included in the decoder module is the same as the number of layers included in the encoder module. Each layer of the decoder module includes an upsampling operation, an attention gate and a residual convolution unit.

[0046] For example, the decoder module contains four layers. The decoder module restores the abstract features to the original resolution through upsampling operations and supplements the feature information of the image by fusing the images with jump connections.

[0047] Figure 6 The structure diagram of the attention gate module includes two input signals, where G is the feature information transmitted by the corresponding encoder's jump connection, and X is the coarse feature output by the previous network layer. The attention gate uses the information extracted from the coarse features to enhance the useful information in the jump connection and suppress irrelevant information and noise.

[0048] In order to solve the problems of low segmentation accuracy and unclear boundary area segmentation that still exist in the prior art in retinal vessel segmentation, the retinal vessel image segmentation method based on differential attention in this embodiment is based on a differential attention multi-scale residual network, which ensures high robustness while obtaining high segmentation accuracy, thereby ensuring accurate retinal fundus vessel image segmentation.

[0049] The following provides the training process of the differential attention multi-scale residual network:

[0050] This embodiment uses the Adam optimizer to optimize network parameters, sets multiple learning rates to optimize the performance of the model, accelerates the convergence of the model during training, and optimizes the performance of the model. In order to meet the training requirements of deep neural networks, we use random data enhancement technology to expand the data set.

[0051] In order to achieve a higher accuracy of retinal blood vessel segmentation to assist in improving the diagnostic efficiency of clinical doctors, the specific implementation method of the proposed differential attention-based multi-scale residual network training will be further described below:

[0052] Step 1: Obtain the public retinal dataset CHASE_DB1, which includes 28 color retinal images of 14 children's eyes, and the pixel size of each image is 999×996; first, grayscale processing is performed on all data to convert RGB images into single-channel grayscale images; then the grayscale images are subjected to adaptive histogram equalization processing.

[0053] Step 2: Considering the memory usage problem during training, the image size is resized to 512×512 pixels and the dataset is randomly divided into training and test sets. In order to meet the training requirements of deep neural networks, the training set is expanded using random data augmentation technology, including random horizontal flipping, random vertical flipping, and random rotation.

[0054] Step 3: Build a multi-scale residual network based on differential attention. The network mainly includes: multi-scale input module, encoder module, differential amplification module, and decoder module. The convolution units in the network are all based on residual connections, as shown in Figure 4, where each convolution block includes a 3×3 convolution layer, a batch normalization layer, and a ReLU activation function layer. The multi-scale input module can retain the original features of the image to a great extent, and to a certain extent avoid the problem of feature loss that occurs in the up and down sampling process of the deep convolutional neural network, making the segmentation results more accurate. The differential amplification module is connected after the residual convolution unit of each encoder to form a jump connection, which is used to extract high-frequency and low-frequency information of images of different resolutions. Figure 5 As shown in the figure, the differential amplification module uses the average pooling operation with a convolution kernel size of 2×2 as a low-pass filter to extract the low-frequency information of the image, and uses the maximum pooling operation with a kernel size of 2×2 as a difference with the convolution operation as a high-pass filter to extract the high-frequency information of the image. Introducing the attention mechanism in the decoder structure can make the network pay more attention to the blood vessel pixels, especially the small blood vessels, effectively suppress the background features and solve the boundary blur problem. Figure 6 As shown in Figure 1, the attention gate has two input signals, one is the feature transmitted by the skip connection, and the other input is the coarse feature output from the previous neural layer. The information extracted from the coarse feature is used to eliminate irrelevant information and noise in the skip connection, making the segmentation result more accurate.

[0055] Among them, the encoder module includes four layers, and the structure of each layer includes two convolutional layers and a pooling layer connected in cascade. Batch normalization operation is performed after each convolutional layer and connected with ReLU activation function. Parameter setting: the convolution kernel size is 3×3, the step size is 1; the pooling layer adopts the maximum pooling operation, the convolution kernel size is 2×2, and the step size is 2. Corresponding to the encoder, the decoder module also includes four layers, each layer includes an upsampling layer and two convolutional layers. Batch normalization operation and ReLU activation function are performed after each convolutional layer to generate output. Parameter setting: the convolution kernel size is 3×3, the step size is 1, and the upsampling receptive field is 2.

[0056] The specific values ​​of the convolution kernel size and stride length in the structure of these modules are fixed. For example, in the multi-scale input module and the differential amplification module, the image resolution size is adjusted by the maximum pooling operation with a 2×2 convolution kernel size and a stride of 2. The convolution kernel size cannot be changed. Usually, non-overlapping pooling operations perform best. The most commonly used size of the convolution kernel for the maximum pooling operation is 2×2. Larger convolution kernels will significantly reduce the size of the output and may cause excessive information loss.

[0057] Step 4: Input the expanded training data set into the network established in step 3 to train the retinal fundus vessel segmentation model. Use the TensorFlow framework to train the model, use the Adam optimizer to optimize the network parameters, set multiple learning rates to optimize the model performance, accelerate the convergence of the model during training, and optimize the performance of the model. Use Tversky Loss to constrain the entire network during training to make the segmentation results more accurate.

[0058] Among them, TensorFlow is a computing framework for deep learning. According to its name, TensorFlow can be divided into tensor + flow. We can understand that within this framework, data is transmitted in the form of tensor flow.

[0059] The Tversky coefficient is a generalization of the Dice coefficient and the Jaccard coefficient and is expressed as:

[0060]

[0061] Where A and B represent the set of predicted values ​​(Prediction) and ground truth (Ground truth), respectively, and α and β are the hyperparameters of the Tversky index. When α=β=0.5, the expression of the Tversky index is the same as the Dice coefficient.

[0062] Among them, the process of optimizing network parameters using Adam optimizer is:

[0063] The model is trained using the TensorFlow framework, and the Adam optimizer is used to optimize network parameters to accelerate the convergence of the model during training and optimize the performance of the model. The Adam algorithm is different from the traditional stochastic gradient descent. The classic stochastic gradient descent method maintains a single learning rate during the training process for all parameter updates, and the learning rate does not change during the training process. The Adam algorithm can calculate the adaptive learning rate of different parameters by calculating the first-order moment estimate and the second-order moment estimate of the gradient: by averaging the past gradient and the current gradient, each gradient update will not differ too much, and can adapt to unstable objective functions; by averaging the past squared gradient and the current squared gradient, an adaptive learning rate can be generated for different parameters.

[0064] Step 5: Load the model parameters saved in step 4 to obtain the trained retinal fundus vessel segmentation model, input the test set selected in step 2 into the model, and obtain the result of retinal fundus vessel segmentation.

[0065] The multi-scale residual network based on differential attention proposed in this embodiment adopts the encoder-decoder structure of U-Net. The differential amplification module is used to improve the jump connection so that the network can learn better features; an attention mechanism is added to each jump connection, and an attention gate is used to highlight the blood vessel pixel part and suppress background features, making the segmentation result more accurate.

[0066] The encoder module of this embodiment compresses the image by downsampling operation, reduces the image resolution and reduces the amount of calculation; the differential amplification module is used to increase the network's extraction of high-frequency information and low-frequency information of images with different resolutions; the attention module allows the network to pay more attention to the required parts while suppressing irrelevant areas in the feature map, making the segmentation result more accurate; the decoder module restores the abstract features to the original resolution through upsampling operation. During the training process, the Tversky loss function is used to constrain the entire network, making the segmentation result more accurate. The Adam optimizer is used to optimize the network parameters and set multiple learning rates to accelerate the convergence of the model during training and optimize the performance of the model.

[0067] Embodiment 2

[0068] This embodiment provides a retinal vascular image segmentation system based on differential attention, which specifically includes:

[0069] An image acquisition module, used for acquiring retinal blood vessel images;

[0070] An image segmentation module is used to obtain a retinal fundus vascular image segmentation result based on a retinal vascular image and a differential attention multi-scale residual network;

[0071] Among them, the differential attention-based multi-scale residual network includes a multi-scale input module, an encoder module, a differential amplification module and a decoder module; the multi-scale input module is used to extract multi-scale information of retinal vascular images; the encoder module is used to encode the multi-scale information; the differential amplification module is used to extract low-frequency information and high-frequency information of the encoded multi-scale information respectively, and then extract the features of the low-frequency information and the high-frequency information respectively; the attention mechanism is introduced in the decoder module to enhance the attention of the required attention area while suppressing irrelevant areas, and finally restore the extracted low-frequency and high-frequency features to the original resolution.

[0072] It should be noted here that the various modules in this embodiment correspond one-to-one to the various steps in Example 1, and the specific implementation process is the same, which will not be repeated here.

[0073] Embodiment 3

[0074] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the retinal vascular image segmentation method based on differential attention as described above.

[0075] Embodiment 4

[0076] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the retinal vascular image segmentation method based on differential attention as described above are implemented.

[0077] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0078] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0079] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0080] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A retinal vascular image segmentation method based on differential attention, characterized in that: include: Acquire retinal vascular images; Based on the retinal vascular image and the differential attention multi-scale residual network, the retinal fundus vascular image segmentation result is obtained; The differential attention-based multi-scale residual network includes a multi-scale input module, an encoder module, a differential amplification module and a decoder module; the multi-scale input module is used to extract multi-scale information of retinal vascular images; the encoder module is used to encode the multi-scale information; the differential amplification module is used to extract low-frequency information and high-frequency information of the encoded multi-scale information, and then extract the features of the low-frequency information and the high-frequency information respectively; the attention mechanism is introduced in the decoder module to enhance the attention of the required attention area while suppressing irrelevant areas, and finally restore the extracted low-frequency and high-frequency features to the original resolution; The differential amplification module is connected after each layer of residual convolution unit of the encoder module to form a skip connection; Each layer of the decoder module includes an upsampling operation, an attention gate and a residual convolution unit; the residual convolution unit includes two convolution blocks, each convolution block includes a 3×3 convolution layer, a batch normalization layer and a ReLU activation function layer; The attention gate includes two input signals, where G is the feature information transmitted by the corresponding jump connection of the encoder, and X is the rough feature output by the previous network layer; the input signal G is sequentially subjected to 3×3 convolution, batch normalization layer, activation function, maximum pooling layer, 3×3 convolution, batch normalization layer and activation function to obtain the intermediate output of G; the input signal X is sequentially subjected to 3×3 convolution, batch normalization layer, activation function, maximum pooling layer, 3×3 convolution, batch normalization layer and activation function to obtain the intermediate output of X; the intermediate output of G and the intermediate output of X perform the addition operation of the corresponding elements of the matrix, and then sequentially pass through 3×3 convolution, batch normalization layer, activation function, transposed convolution and sigmoid activation function to obtain the main output, and the main output is multiplied with the input signal X to obtain the final output.

2. The retinal vascular image segmentation method based on differential attention according to claim 1, characterized in that: The multi-scale input module includes inputs of at least three scales, the input of the first scale is a preprocessed grayscale image of the original training set, and continuous maximum pooling operations are performed on the input of the first scale to obtain inputs of the second scale and inputs of scales after the second scale in turn.

3. The retinal vascular image segmentation method based on differential attention according to claim 1, characterized in that: The encoder module contains the same number of layers as the number of input scales of the multi-scale input module.

4. The retinal vascular image segmentation method based on differential attention as claimed in claim 3, characterized in that: Each layer of the encoder module includes a residual convolution unit and a downsampling operation. The residual convolution unit is used to fuse multi-scale input images to supplement feature information and compress the image to reduce the image resolution through the downsampling operation.

5. The retinal vascular image segmentation method based on differential attention as claimed in claim 1, characterized in that: In the differential amplification module, the average pooling operation is used as a low-pass filter to extract low-frequency information, and the difference between the maximum pooling operation and the convolution operation is used as a high-pass filter to extract high-frequency information. The high-frequency information and the low-frequency information are convolved separately to further extract features.

6. The retinal vascular image segmentation method based on differential attention according to claim 1, characterized in that: The number of layers included in the decoder module is the same as the number of layers included in the encoder module.

7. A retinal vascular image segmentation system based on differential attention, characterized in that: include: An image acquisition module, used for acquiring retinal blood vessel images; An image segmentation module is used to obtain a retinal fundus vascular image segmentation result based on a retinal vascular image and a differential attention multi-scale residual network; The differential attention-based multi-scale residual network includes a multi-scale input module, an encoder module, a differential amplification module and a decoder module; the multi-scale input module is used to extract multi-scale information of retinal vascular images; the encoder module is used to encode the multi-scale information; the differential amplification module is used to extract low-frequency information and high-frequency information of the encoded multi-scale information, and then extract the features of the low-frequency information and the high-frequency information respectively; the attention mechanism is introduced in the decoder module to enhance the attention of the required attention area while suppressing irrelevant areas, and finally restore the extracted low-frequency and high-frequency features to the original resolution; The differential amplification module is connected after each layer of residual convolution unit of the encoder module to form a skip connection; Each layer of the decoder module includes an upsampling operation, an attention gate and a residual convolution unit; the residual convolution unit includes two convolution blocks, each convolution block includes a 3×3 convolution layer, a batch normalization layer and a ReLU activation function layer; The attention gate includes two input signals, where G is the feature information transmitted by the corresponding jump connection of the encoder, and X is the rough feature output by the previous network layer; the input signal G is sequentially subjected to 3×3 convolution, batch normalization layer, activation function, maximum pooling layer, 3×3 convolution, batch normalization layer and activation function to obtain the intermediate output of G; the input signal X is sequentially subjected to 3×3 convolution, batch normalization layer, activation function, maximum pooling layer, 3×3 convolution, batch normalization layer and activation function to obtain the intermediate output of X; the intermediate output of G and the intermediate output of X perform the addition operation of the corresponding elements of the matrix, and then sequentially pass through 3×3 convolution, batch normalization layer, activation function, transposed convolution and sigmoid activation function to obtain the main output, and the main output is multiplied with the input signal X to obtain the final output.

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

9. A computer 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 program, the steps in the retinal vascular image segmentation method based on differential attention as described in any one of claims 1 to 6 are implemented.

Images