A segmentation method for diabetic retinopathy based on Poisson fusion data augmentation

By constructing the U-Net++ network and combining residual modules, lightweight channel attention mechanisms, and depthwise separable convolutions, the problem of inaccurate information extraction in the segmentation of diabetic retinopathy by convolutional neural networks was solved, achieving higher segmentation accuracy and efficiency.

CN118799336BActive Publication Date: 2026-11-17SUZHOU UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410827232.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-25
Publication Date
2026-11-17
Estimated Expiration
2044-06-25

AI Technical Summary

Technical Problem

Existing convolutional neural networks suffer from problems in image segmentation of diabetic retinopathy, such as the inability to accurately extract effective information and the extraction of redundant information, resulting in large errors in the output results.

Method used

We employ a Poisson-based data augmentation method by constructing a U-Net++ network. This network utilizes residual modules and a lightweight channel attention mechanism, combined with depthwise separable convolutions and a lightweight ASPP module, to enhance information extraction and suppress redundant information, thereby improving the performance of the segmentation network.

Benefits of technology

It improves the accuracy and efficiency of the model in segmenting diabetic retinopathy, reduces computational cost and number of parameters, enhances the utilization of image context features, and improves segmentation performance in complex tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118799336B_ABST
    Figure CN118799336B_ABST
Patent Text Reader

Abstract

This invention discloses a segmentation method for diabetic retinopathy based on Poisson fusion data augmentation, belonging to the field of medical image segmentation technology. The invention includes: receiving lesion images; augmenting the lesion images using a Poisson fusion algorithm based on mean coordinates; preprocessing the images; and then dividing the dataset into training and testing sets. A segmentation network structure based on U-Net++ is constructed, including an encoder and a decoder. In the encoder, standard convolutional modules are replaced with residual modules, and a lightweight channel attention mechanism is added after each residual module. This invention adds residual modules and a lightweight channel attention mechanism between the encoder and decoder, which not only helps preserve the original input information and maintain information flow as network depth increases, allowing the model to better focus on important parts of the input data and avoid information loss, but also effectively improves model performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation technology, specifically to a segmentation method for diabetic retinopathy based on Poisson fusion data augmentation. Background Technology

[0002] Diabetic retinopathy (DR) is a complication of diabetes that poses a significant risk to the vision of the global working-age population. According to the latest data released by the International Diabetes Federation (IDF), one in three people with diabetes will develop DR, and one in three DR patients will face the risk of blindness. Early diagnosis can effectively prevent the progression of DR. However, DR imposes a huge economic burden on society each year, and DR screening places a heavy burden on ophthalmologists. Therefore, computer-assisted diagnosis of the disease is a significant challenge and is crucial.

[0003] In existing technologies, convolutional neural networks are often used to extract image features during image segmentation. However, since convolutional neural networks typically set the number of feature map channels to be very large, the encoder may fail to extract effective information and extract useless information, resulting in errors in the output results. To address this, we propose a segmentation method for diabetic retinopathy based on Poisson fusion data augmentation. Summary of the Invention

[0004] The purpose of this invention is to provide a segmentation method for diabetic retinopathy based on Poisson fusion data augmentation, which can enhance effective information, suppress redundant information, and solve the problem of inaccurate extraction of effective information.

[0005] According to a first aspect of the present invention, in order to achieve the above-mentioned objective, the present invention provides the following technical solution: a method for segmenting diabetic retinopathy based on Poisson fusion data augmentation, comprising the following steps:

[0006] The lesion images are received, and the lesion images are augmented using a Poisson fusion algorithm based on mean coordinates. The images are then preprocessed, and the dataset is divided into a training set and a test set.

[0007] A segmentation network structure based on U-Net++ is constructed, which includes an encoder and a decoder. In the encoder, the standard convolutional module is replaced with a residual module, and a lightweight channel attention mechanism is added after each residual module.

[0008] A lightweight ASPP module is added to the bottom layer of the segmentation network structure, and the original standard convolutions in the segmentation network structure are replaced with depthwise separable convolutions.

[0009] The above segmentation network structure is trained using the training set to obtain the trained segmentation network model;

[0010] The test set images are input into the segmentation network model described above to obtain the prediction result image, and the prediction result is evaluated.

[0011] Furthermore, the lesion images are downloaded using the e-ophtha dataset.

[0012] Furthermore, a Poisson fusion algorithm based on mean coordinates is used to augment the lesion image, as follows:

[0013] (1) Determine the target image and background image;

[0014] (2) Set the location of the background area to be merged in the target image;

[0015] (3) Fill in the A matrix for solving Poisson;

[0016] (4) Calculate and solve for the B matrix of Poisson;

[0017] a. The Laplacian operator directly calculates the divergence of the target image to obtain the initial B matrix;

[0018] b. Fill the boundary pixels of matrix B according to the boundary conditions;

[0019] (5) Use the least squares method to solve Ax = B. Since A and B are known, x can be obtained directly, which is the fused image.

[0020] Furthermore, the lesion images are preprocessed, specifically including cropping the images to the same pixel size and flipping, rotating, and affine transforming the images.

[0021] Furthermore, the feature extractor of the encoder comprises five stages, each consisting of a residual module and a lightweight channel attention mechanism. Each residual module contains a depthwise separable convolutional layer, a normalization layer, and an activation function layer. Max pooling layers are used between each stage to reduce the feature map and increase the receptive field of the convolution.

[0022] Furthermore, the decoder uses the decoding part of UNet++ to reconstruct high-resolution feature maps, which includes 5 stages. Each stage consists of two depthwise separable convolutional modules. Each module contains a depthwise separable convolutional layer, a normalization layer, and an activation function layer. An upsampling method is used between each stage to increase the feature map and gradually restore the image resolution.

[0023] Furthermore, the lightweight ASPP module is based on the traditional ASPP and has been improved in the following ways:

[0024] (1) The original global pooling branch was removed, and the original four convolution branches were reduced to three. The dilation rates of these three dilated convolution branches were selected as 1, 3, and 5, respectively.

[0025] (2) The original direct merging branch is improved into a skip connection, which enables the network to capture rich contextual semantic information.

[0026] Furthermore, the depthwise separable convolution is expressed as follows:

[0027]

[0028] In the formula, K′ represents the filter of the depthwise convolution, G′ represents the output feature map of the depthwise convolution, M represents the input image, i represents the stride in the horizontal direction, j represents the stride in the vertical direction, k represents the number of convolution kernels or channels, m represents the height of the output feature map, and l represents the depth of the input feature map.

[0029] According to a second aspect of the present invention, the present invention provides a segmentation system for diabetic retinopathy based on Poisson fusion data augmentation, comprising:

[0030] The preprocessing unit receives lesion images, augments them using a Poisson fusion algorithm based on mean coordinates, preprocesses them, and then divides the dataset into training and testing sets.

[0031] The module building unit is used to build a segmentation network structure based on the U-Net++ network. The segmentation network structure includes an encoder and a decoder. In the encoder, the standard convolutional module is replaced with a residual module, and a lightweight channel attention mechanism is added after each residual module.

[0032] Convolutional pooling units are used to add lightweight ASPP modules to the bottom layer of the segmentation network structure and replace the original standard convolutions in the segmentation network structure with depthwise separable convolutions.

[0033] The training unit is used to train the above segmentation network structure using the training set to obtain the trained segmentation network model.

[0034] The prediction unit is used to input the test set images into the segmentation network model to obtain the prediction result map and evaluate the prediction result.

[0035] According to a third aspect of the present invention, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor, and when the processor loads and executes the computer program, it employs the above-described diabetic retinopathy segmentation method based on Poisson fusion data augmentation.

[0036] The present invention has at least the following beneficial effects:

[0037] 1. This invention adds a residual module and a lightweight channel attention mechanism between the encoder and decoder, which not only helps to preserve the original input information and maintain the flow of information as the network depth increases, but also enables the model to better focus on the important parts of the input data, avoid the problem of information loss, and effectively improve the performance of the model.

[0038] 2. This invention enhances the receptive field of the feature extraction network by adding a lightweight ASPP module, enabling the segmentation network to more fully utilize the feature information at different levels of the image context and effectively fuse them, thus helping the segmentation network to exhibit higher accuracy in complex tasks.

[0039] 3. This invention enables the model to significantly reduce the number of parameters and computational cost while maintaining performance through a depthwise separable convolution model, thereby improving segmentation accuracy and efficiency.

[0040] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the segmentation network structure in this invention;

[0042] Figure 2 This is a schematic diagram of the residual module and the lightweight channel attention mechanism in this invention;

[0043] Figure 3 This is a schematic diagram of the structure of the traditional ASPP module and the lightweight ASPP module in this invention;

[0044] Figure 4 This is a schematic diagram of the depth-separable convolution structure in this invention;

[0045] Figure 5 This is a comparative diagram of the visual segmentation results of the ablation experiment in this invention. Detailed Implementation

[0046] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0047] Example 1:

[0048] Please see Figure 1-5 This invention provides a technical solution: a segmentation method for diabetic retinopathy based on Poisson fusion data augmentation, comprising the following steps:

[0049] S1. Receive lesion images, augment the lesion images using a Poisson fusion algorithm based on mean coordinates, preprocess the images, and then divide the dataset into training and test sets.

[0050] It should be noted that the data in this embodiment comes from a color fundus image database specifically for research on diabetic retinopathy (DR), specifically the e-ophtha dataset;

[0051] Furthermore, a Poisson fusion algorithm based on mean coordinates is used to augment the lesion image, as follows:

[0052] (1) Determine the target image and background image;

[0053] (2) Set the location of the background area to be merged in the target image;

[0054] (3) Fill in the A matrix for solving Poisson;

[0055] (4) Calculate and solve for the B matrix of Poisson;

[0056] a. The Laplacian operator directly calculates the divergence of the target image to obtain the initial B matrix;

[0057] b. Fill the boundary pixels of matrix B according to the boundary conditions;

[0058] (5) Use the least squares method to solve Ax = B. Since A and B are known, x can be obtained directly, which is the fused image.

[0059] Furthermore, the lesion images are preprocessed, specifically including cropping the images to the same pixel size and flipping, rotating, and affine the images;

[0060] Regarding the technical solution of this embodiment, the EX lesion subset includes 82 images, divided into 47 images with EX lesions and 35 normal images. These images have varying resolutions, ranging from 1504×1000 pixels to 2544×1696 pixels. In this embodiment, 47 images containing EX lesions were selected from the e-ophtha dataset. These images were divided into training and testing sets, with 32 images used for training and the remaining 15 images used for testing. To facilitate training and reduce memory consumption, these images were cropped and adjusted to a uniform size of 1024×1024 pixels. In addition, this invention also performed transformations on the images such as flipping, affine transformation, and rotation that do not affect the authenticity of the data. Through these methods, the generalization ability of the learning model is significantly enhanced, the scale of the training dataset is expanded, and the complexity and variability of the data are also improved.

[0061] S2. Construct a segmentation network structure based on the U-Net++ network. The segmentation network structure includes an encoder and a decoder. In the encoder, the standard convolutional module is replaced with a residual module, and a lightweight channel attention mechanism is added after each residual module.

[0062] Specifically, the encoder’s feature extractor consists of five stages, each consisting of a residual module and a lightweight channel attention mechanism. Each residual module contains a depthwise separable convolutional layer, a normalization layer and an activation function layer. Max pooling layers are used between each stage to reduce the feature map and increase the receptive field of the convolution.

[0063] The decoder uses the decoding part of UNet++ to reconstruct high-resolution feature maps. It consists of 5 stages, each stage is composed of two depthwise separable convolutional modules. Each module contains a depthwise separable convolutional layer, a normalization layer and an activation function layer. An upsampling method is used between each stage to increase the feature map and gradually restore the resolution of the image.

[0064] like Figure 2 The diagram shows the DRC module, which consists of DSR-Part and LWCAM. DSR-Part is a residual module, and each module contains depthwise separable convolutional layers, normalization layers, and activation function layers. LWCAM is a lightweight channel attention mechanism, which is added to the end of each DRC-Part.

[0065] The LWCAM proposed in this embodiment performs an Add operation after global max pooling and global average pooling, and then processes the result through MLP to reduce and increase the dimensionality, thereby obtaining the weight matrix.

[0066] The residual module and lightweight channel attention mechanism are used to maintain the flow of information. Since convolutional neural networks usually set the number of feature map channels to be very large, the encoder may not be able to extract effective information well, and the extraction of useless information may mislead the network output. Therefore, this part can enhance effective information, suppress redundant information, help the model better focus on the important parts of the input data, avoid the problem of information loss, and effectively improve the performance of the model.

[0067] S3. Add a lightweight ASPP module to the bottom layer of the segmentation network structure, and replace the original standard convolutions in the segmentation network structure with depthwise separable convolutions;

[0068] Specifically, the lightweight ASPP module is used to enhance the receptive field of the segmentation network, enabling the segmentation network to make fuller use of the feature information at different levels of the image context and effectively fuse them, which helps the segmentation network to show higher accuracy in complex tasks.

[0069] Regarding the technical solution of this embodiment, such as Figure 3 As shown, the lightweight ASPP module is used to fuse and extract contextual semantic information. This module mainly makes the following improvements based on the traditional ASPP:

[0070] (1) The original global pooling branch was removed, and the original four convolution branches were reduced to three. The dilation rates of these three dilated convolution branches were selected as 1, 3 and 5, respectively, in order to capture contextual information at different scales while maintaining computational efficiency. Through this improvement, the optimized ASPP module reduces the consumption of computing resources and improves processing speed while maintaining performance;

[0071] (2) The original direct merging branch is improved to a skip connection. This improvement enables the network to capture richer contextual semantic information, thereby further improving the network's performance. Through skip connections, feature information from different levels can be effectively fused, which helps the network to show higher accuracy in complex tasks.

[0072] Furthermore, such as Figure 4 As shown, depthwise separable convolution replaces the original standard convolution module; standard convolution operations involve a large number of parameters and have high computational costs, which is particularly prominent in resource-constrained scenarios. To solve this problem, this embodiment uses depthwise separable convolution, which reduces the number of parameters and computational costs while maintaining performance.

[0073] like Figure 4 As shown, a set of dimensions D M ×D M The feature map M of size D is used as input, and a set of features of size D is used as input. G ×DG The feature map G of ×N is used as the output, D M Let M be the width and height of the input feature map, M be the number of input image channels, N be the number of output image channels, and D be the width and height of the input feature map. K ×D K The kernel sizes for standard convolution and depthwise convolution are given. The kernel size for pointwise convolution is 1×1 with a stride of 1, and the standard convolution process with edge addition is shown in equation (1):

[0074]

[0075] In the formula, K′ represents the filter of the depthwise convolution, G′ represents the output feature map of the depthwise convolution, M represents the input image, i represents the stride in the horizontal direction, j represents the stride in the vertical direction, k represents the number of convolution kernels or channels, m represents the height of the output feature map, n represents the width of the output feature map, and l represents the depth of the input feature map. The computational cost of standard convolution is D. M 2 ×M×N×D k 2 ;

[0076] The depthwise convolution operation for one filter per input channel is shown in equation (2):

[0077]

[0078] In the formula, K′ represents the filter of the depthwise convolution, G′ represents the output feature map of the depthwise convolution, M represents the input image, i represents the stride in the horizontal direction, j represents the stride in the vertical direction, k represents the number of convolution kernels or channels, m represents the height of the output feature map, and l represents the depth of the input feature map.

[0079] The m-th filter in K′ is applied to the m-th feature map in M ​​to produce the m-th feature map in G′. The computational complexity of the depthwise convolution is D. M 2 ×M×1×D k 2 The computational cost of pointwise convolution is D. M 2 Therefore, the computational cost of depthwise separable convolution is D × 1 × 1 × M × N. M 2 ×M×D k 2 +1×1×M×N×D M 2 The computational complexity of standard convolution and depthwise separable convolution is shown in equation (3):

[0080]

[0081] The pattern revealed by equation (3) is that as the number of output feature maps N increases, depthwise separable convolution not only reduces the number of required parameters, but also reduces computational complexity. This means that the number of kernel moves is reduced during model training and inference. This feature enables depthwise separable convolution to effectively improve the performance and efficiency of the model.

[0082] S4. Train the above segmentation network structure using the training set to obtain the trained segmentation network model;

[0083] S5. Input the test set images into the above segmentation network model to obtain the prediction result image, and evaluate the prediction result.

[0084] Specifically, in order to quantitatively evaluate the performance of the method proposed in this invention, the following four evaluation metrics commonly used in medical image segmentation are used as the criteria for measuring the quality of experimental results: Dice coefficient (DSC), recall, precision, and intersection-union ratio (IOU).

[0085] DSC is a function that evaluates the similarity between predicted results and manually labeled data. It is obtained by using the ratio of the intersection to the union of the predicted results and manually labeled data, and is defined as:

[0086]

[0087] Where TP (True Positive) represents the number of true positive pixels, FP (False Positive) represents the number of false positive pixels, and FN (False Negative) represents the number of false negative pixels.

[0088] Recall represents the proportion of correctly identified pixels out of all manually labeled pixels, precision represents the proportion of correctly identified pixels out of all predicted pixels, and the Intersection over Union (IOU) coefficient is used to evaluate the degree of overlap between the lesion region predicted by the model and the actual lesion region. Recall, precision, and IOU are calculated as follows:

[0089]

[0090] The precision-recall curve (PR curve) and the corresponding average precision-recall (AUPR) play an important role in evaluating the performance of diabetic retinopathy (DR) lesion segmentation. In this embodiment, the AUPR curve is generated by applying threshold processing to the results of 33 equally spaced instances. This method ensures that the evaluation method of the present invention can more accurately measure the performance of the model in the DR lesion segmentation task.

[0091] The method proposed in this embodiment is compared with U-Net, U-Net++, U-Net3+, Deeplab-v3+ and DSR-U-Net++. The corresponding quantitative segmentation results and ablation experiment results are shown in Table 1. AS is the baseline model DSR-U-Net++, and the last three rows are ablation experiments. The results show that the three modules proposed in this embodiment are effective.

[0092] AS+DRC+LWASPP+DSConv represents the method of this embodiment, which adds a residual module, a lightweight attention mechanism, a lightweight ASPP module, and depthwise separable convolution to DSR-U-Net++. Among these methods, U-Net++, U-Net3+, and DSR-U-Net++ have better AUPR than U-Net. Compared with other methods, Deeplab-v3+ generally has lower performance, indicating that U-shaped models are more advantageous in lesion segmentation tasks. Compared with these five baseline models, the method proposed in this embodiment shows excellent performance in the detection of EX lesions. The segmentation results show that in terms of EX lesion segmentation, this embodiment shows the best performance in both preserving the integrity of large lesions and detecting small lesions.

[0093] Table 1. Quantitative segmentation results and ablation experimental results of different methods

[0094]

[0095]

[0096] Example 2:

[0097] This invention provides a segmentation system for diabetic retinopathy based on Poisson fusion data augmentation, characterized in that it includes:

[0098] The preprocessing unit receives lesion images, augments them using a Poisson fusion algorithm based on mean coordinates, preprocesses them, and then divides the dataset into training and testing sets.

[0099] The module building unit is used to build a segmentation network structure based on the U-Net++ network. The segmentation network structure includes an encoder and a decoder. In the encoder, the standard convolutional module is replaced with a residual module, and a lightweight channel attention mechanism is added after each residual module.

[0100] Convolutional pooling units are used to add lightweight ASPP modules to the bottom layer of the segmentation network structure and replace the original standard convolutions in the segmentation network structure with depthwise separable convolutions.

[0101] The training unit is used to train the above segmentation network structure using the training set to obtain the trained segmentation network model.

[0102] The prediction unit is used to input the test set images into the segmentation network model to obtain the prediction result map and evaluate the prediction result.

[0103] Specifically, the aforementioned preprocessing unit, module construction unit, convolutional pooling unit, training unit, and prediction unit can be embedded in a computer processing system. The computer, based on the aforementioned diabetic retinopathy segmentation method based on Poisson fusion data augmentation, calls upon each of the aforementioned units to complete the task of equivalent modeling. The aforementioned preprocessing unit, module construction unit, convolutional pooling unit, training unit, and prediction unit can perform operations according to the specific steps given in the aforementioned diabetic retinopathy segmentation method based on Poisson fusion data augmentation.

[0104] It should be noted that the division of the various units in the above system is merely a division of logical functions. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. These units can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some units can be implemented by processing element calls to software, while others are implemented in hardware. For example, the receiving unit can be a separate processing element, or it can be integrated into a chip in the aforementioned device. Alternatively, it can be stored as program code in the memory of the aforementioned device, and its signal processing unit functions can be called and executed by a processing element of the device. The implementation of other units is similar. Furthermore, these units can be fully or partially integrated together, or they can be implemented independently. The processing element mentioned here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above units can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0105] For example, these units can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when one of the above units is implemented in the form of processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these units can be integrated together to implement a system-on-a-chip (SOC).

[0106] Example 3:

[0107] The present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor. When the processor loads and executes the computer program, it employs the above-mentioned diabetic retinopathy segmentation method based on Poisson fusion data augmentation.

[0108] It should be noted that the terminal device can be a computer device such as a desktop computer, a laptop computer, or a cloud server, and the terminal device includes, but is not limited to, a processor and a memory. For example, the terminal device may also include input / output devices, network access devices, and buses.

[0109] Furthermore, the processor can be a central processing unit (CPU). Of course, depending on the actual use, other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be used. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it in this regard.

[0110] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0111] For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances. When an element is referred to as being "assembled on," "mounted on," "fixed to," or "set on" another element, it may be directly on the other element or there may be an intermediate element present. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be an intermediate element present. The terms "vertical," "horizontal," "upper," "lower," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible embodiments.

[0112] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0113] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

Claims

1. A method for diabetic retinopathy segmentation based on Poisson fusion data augmentation, characterized in that, Includes the following steps: The lesion images are received, and the lesion images are augmented using a Poisson fusion algorithm based on mean coordinates. The images are then preprocessed, and the dataset is divided into a training set and a test set. A segmentation network structure based on U-Net++ is constructed, which includes an encoder and a decoder. In the encoder, the standard convolutional module is replaced with a residual module, and a lightweight channel attention mechanism is added after each residual module. The encoder’s feature extractor consists of five stages, each consisting of a residual module and a lightweight channel attention mechanism. Each residual module contains a depthwise separable convolutional layer, a normalization layer and an activation function layer. Max pooling layers are used between each stage to reduce the feature map and increase the receptive field of the convolution. The decoder uses the decoding part of UNet++ to reconstruct high-resolution feature maps. It consists of 5 stages, each stage is composed of two depthwise separable convolutional modules. Each module contains a depthwise separable convolutional layer, a normalization layer and an activation function layer. An upsampling method is used between each stage to increase the feature map and gradually restore the resolution of the image. A lightweight ASPP module is added to the bottom layer of the segmentation network structure, and the original standard convolutions in the segmentation network structure are replaced with depthwise separable convolutions. The lightweight ASPP module is based on the traditional ASPP and has the following improvements: the original global pooling branch has been removed, and the original four convolution branches have been reduced to three. The dilation rates of these three dilated convolution branches are selected as 1, 3 and 5 respectively. The original direct merging branch has been improved into a skip connection, which enables the network to capture rich contextual semantic information. The above segmentation network structure is trained using the training set to obtain the trained segmentation network model; The test set images are input into the segmentation network model described above to obtain the prediction result image, and the prediction result is evaluated.

2. The method of claim 1, wherein the method is based on Poisson fusion data augmentation for diabetic retinopathy segmentation. The lesion images were downloaded using the e-ophtha dataset.

3. The method of claim 2, wherein the method is based on Poisson fusion data augmentation for diabetic retinopathy segmentation. The lesion image is augmented using a Poisson fusion algorithm based on mean coordinates, as detailed below: (1) Determine the target image and background image; (2) Set the location of the background area to be merged in the target image; (3) Fill in the A matrix to solve for Poisson; (4) Calculate and solve for the B matrix of Poisson; a. The Laplacian operator directly calculates the divergence of the target image to obtain the initial B matrix; b. Fill the boundary pixels of matrix B according to the boundary conditions; (5) Use the least squares method to solve Ax=B. Since A and B are known, x can be obtained directly, that is, the fused image.

4. The method of claim 2, wherein the method is based on Poisson fusion data augmentation for diabetic retinopathy segmentation, characterized by, Preprocessing of lesion images specifically includes cropping the images to the same pixel size and performing flipping, rotation, and affine transformation on the images.

5. The method of claim 4, wherein the method is based on Poisson fusion data augmentation for diabetic retinopathy segmentation. The depthwise separable convolution is expressed as follows: wherein a filter representing a depthwise convolution, an output feature map representing a depthwise convolution, M represents an input image, i represents a step in a horizontal direction, j represents a step in a vertical direction, k represents a number of convolution kernels or a number of channels, m represents a height of an output feature map, and l represents a depth of an input feature map.

6. A Poisson fusion data augmentation based diabetic retinopathy segmentation system for implementing the Poisson fusion data augmentation based diabetic retinopathy segmentation method of any one of claims 1 to 5, characterized in that, include: The preprocessing unit receives lesion images, augments them using a Poisson fusion algorithm based on mean coordinates, preprocesses them, and then divides the dataset into training and testing sets. The module building unit is used to build a segmentation network structure based on the U-Net++ network. The segmentation network structure includes an encoder and a decoder. In the encoder, the standard convolutional module is replaced with a residual module, and a lightweight channel attention mechanism is added after each residual module. Convolutional pooling units are used to add lightweight ASPP modules to the bottom layer of the segmentation network structure and replace the original standard convolutions in the segmentation network structure with depthwise separable convolutions. The training unit is used to train the above segmentation network structure using the training set to obtain the trained segmentation network model. The prediction unit is used to input the test set images into the segmentation network model to obtain the prediction result map and evaluate the prediction result.

7. A terminal device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on a processor. When the processor loads and executes the computer program, it employs the diabetic retinopathy segmentation method based on Poisson fusion data augmentation as described in any one of claims 1 to 5.

Citation Information

Patent Citations

  • Improved U-Net brain tumor segmentation method based on attention mechanism and multi-scale feature fusion

    CN115424103A

  • Paravascular anomaly segmentation method based on double decoders and local feature enhancement network

    CN118172369A