A Diabetic Retinopathy Assistant Diagnosis System

Abstract

The present invention relates to the field of medical diagnosis technology, and a diabetic retinopathy auxiliary diagnosis system, which includes a data acquisition and upload module for acquiring fundus images of patients; a diabetic retinopathy grading module for grading diabetic retinopathy of the fundus images; a diabetic retinopathy segmentation module for inputting the fundus images into the diabetic retinopathy segmentation module for lesion segmentation; an intelligent auxiliary diagnosis and analysis module for automatically analyzing information such as the type, quantity, size, etc. of diabetic retinopathy, displaying it to the user together with the diabetic retinopathy grading result, and finally generating a graphic report to assist ophthalmologists in better diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of medical diagnosis, and particularly to an auxiliary diagnosis system for diabetic retinopathy. Background Art

[0002] Diabetes is a metabolic disease with long-term hyperglycemia, and the incidence of diabetes is very high, with a large population worldwide. Diabetic retinopathy is a complication of diabetes, and more than one-third of diabetic patients are affected by diabetic retinopathy. Moreover, diabetic retinopathy is also the main cause of blindness in adults, and has become one of the major diseases threatening human health.

[0003] At present, the auxiliary diagnosis methods for diabetic retinopathy are divided into two types: traditional machine learning methods and deep learning algorithms. Most of the traditional machine learning algorithms achieve segmentation from a single morphological feature. This method relies on professional medical knowledge, and the cost of obtaining medical data is relatively high compared to ordinary data. Therefore, the algorithm accuracy and generalization ability of traditional machine learning methods are relatively low compared to deep learning algorithms. Applying convolutional neural networks to the segmentation task can solve many problems existing in traditional machine learning to a large extent and greatly improve the algorithm accuracy and generalization ability. However, in specific segmentation tasks, convolutional neural networks still have problems such as being unable to distinguish similar feature regions and unable to segment regions with small volume and blurred edges. This results in low accuracy and weak generalization ability of the method. Existing auxiliary diagnosis systems have problems such as slow processing speed and inability to perform real-time detection. Therefore, there is still great room for improvement in the field of auxiliary diagnosis systems.

[0004] Therefore, the research on developing an intelligent auxiliary diagnosis system for diabetic retinopathy has very important clinical significance. Summary of the Invention

[0005] To solve the deficiencies in the prior art, the present invention provides an auxiliary diagnosis system for diabetic retinopathy based on deep learning, which can not only detect minor lesions but also meet actual needs, providing reasonable diagnostic basis for ophthalmologists. The present invention adopts the following technical solutions:

[0006] The technical solution adopted by the present invention is: an auxiliary diagnosis system for diabetic retinopathy, comprising

[0007] a data acquisition and upload module, configured to acquire fundus images of patients and upload them to a diabetic retinopathy grading module and a diabetic retinopathy segmentation module;

[0008] The diabetic retinopathy grading module is used to grade the received fundus images for diabetic retinopathy, obtain the diabetic retinopathy grade of the fundus images, and transmit it to the intelligent assisted diagnosis and analysis module;

[0009] The diabetic retinopathy segmentation module is used to input the received fundus images into the diabetic retinopathy segmentation module for lesion segmentation, obtain the lesion segmentation map of the fundus images, and transmit it to the intelligent assisted diagnosis and analysis module;

[0010] The intelligent assisted diagnosis and analysis module is used to comprehensively analyze the received lesion segmentation map of the fundus images, and generate a graphic report in combination with the received diabetic retinopathy grade of the fundus images, and display it to the user.

[0011] As a preferred mode: The diabetic retinopathy includes hemorrhage points, hard exudates, soft exudates, microaneurysms, and optic discs.

[0012] As a preferred mode: The comprehensive analysis includes analyzing information such as the type, location, quantity, quantity, and contour of diabetic retinopathy.

[0013] As a preferred mode: The specific diagnosis method of the system includes the following steps:

[0014] S1. The data acquisition and upload module acquires the original fundus images and performs preprocessing to obtain preprocessed fundus images, and then uploads the preprocessed fundus images to the diabetic retinopathy grading module and the diabetic retinopathy segmentation module;

[0015] S2. After receiving the preprocessed fundus images, the diabetic retinopathy grading module uses the pre-trained diabetic retinopathy grading model to grade the preprocessed fundus images for diabetic retinopathy, obtain the diabetic retinopathy grade of the fundus images, and transmit it to the intelligent assisted diagnosis and analysis module;

[0016] S3. After receiving the preprocessed fundus images, the diabetic retinopathy segmentation module uses the pre-trained diabetic retinopathy segmentation model to perform image segmentation on the preprocessed fundus images, obtain the lesion segmentation map of the fundus images, and transmit it to the intelligent assisted diagnosis and analysis module;

[0017] S4. After receiving the lesion segmentation map of the fundus images, the intelligent assisted diagnosis and analysis module performs comprehensive analysis to obtain the contour and location information of the lesions in the fundus images, and then generates a graphic report in combination with the received diabetic retinopathy grade of the fundus images, and displays it to the user.

[0018] As a preferred method: The original fundus image described in step S1 can be obtained by photographing the patient's eye with a fundus camera, which can be a real-time captured image or an image stored in a local server or the cloud, and is obtained by means of reading or network transmission.

[0019] As a preferred method: The diabetic retinopathy grading model described in step S2 adopts a GhostNet network model and grades the preprocessed fundus image according to the grading standard; when training the GhostNet network model, the existing graded fundus image dataset is divided into a training set, a validation set and a test set, and then the training set is input into the GhostNet network model for training, the validation set is used for auxiliary verification training, and the test set is used for detection to obtain a trained diabetic retinopathy grading model with a detection accuracy reaching the set requirements.

[0020] As a preferred method: The GhostNet network model is composed of 1 Conv3, 16 G-bneck modules, 1 Conv1, 1 Pooling and 1 FC. The Conv3 represents a convolutional layer with a convolutional kernel size of 3×3 and a stride of 2; the Conv1 represents a convolutional layer with a convolutional kernel size of 1×1; the Pooling is an average pooling layer; the FC is a fully connected layer; the G-bneck module is divided into a G-bneck1 module and a G-bneck2 module. The G-bneck1 module consists of 2 Ghost structures, 2 batch normalization layers and 1 ReLU activation function. The G-bneck2 module consists of 2 Ghost structures, a Depthwise convolutional layer with a stride of 2, 3 batch normalization layers and 2 ReLU activation functions. During training, the SGD optimizer is adopted, and the loss function adopts the cross-entropy loss function.

[0021] As a preferred method: The diabetic retinopathy segmentation model adopts a DRSU-Net (Diabetesretinopathy segmentation U-Net) network model; when training the DRSU-Net network model, the existing segmented fundus image lesion segmentation map dataset is divided into a training set, a validation set and a test set, and then the training set is input into the DRSU-Nett network model for training, the validation set is used for auxiliary verification training, and the test set is used for detection to obtain a trained diabetic retinopathy segmentation model with a detection accuracy reaching the set requirements. During training, the Adam optimizer is adopted, and the loss function adopts the binary cross-entropy loss function.

[0022] As a preferred method: the DRSU-Net network model includes a convolutional layer, a pooling layer, a deconvolutional layer, and an activation function; the pooling layer uses a max pooling operation; the activation function uses a ReLU linear rectification unit; the convolutional layer is used to extract high-dimensional feature information of the image, and the deconvolutional layer is used to restore the feature map to the size of the original input image.

[0023] As a preferred method: the DRSU-Net network model includes an encoder part, a decoder part, a bridging part, and a skip connection part. The encoder part is used to obtain high-level abstract features of the image; the decoder part is used to restore the feature map to its original size; the bridging part is used to connect the encoder part and the decoder part; the skip connection part is used to fuse the feature information of the encoder part and the decoder part.

[0024] The beneficial effects of the present invention are: the system of the present invention can not only determine the diabetic retinopathy grade of fundus images, but also mark the position and contour of diabetic retinopathy in fundus images. And the algorithm of the lesion segmentation part of this system can effectively solve the problem of poor detection effect for tiny lesions.

[0025] The present invention conducts auxiliary diagnosis of diabetic retinopathy for fundus images, which can improve the diagnostic accuracy and efficiency of diabetic retinopathy, and at the same time can reduce misdiagnosis and missed diagnosis caused by fatigue and other situations. Specific Embodiments

[0026] The technical solution of the present invention will be described specifically and in detail below so that the technical features of the present invention can be more easily understood by those skilled in the art. It should be noted that the specific embodiments listed here are only illustrative descriptions of the present invention, rather than limitations on the protection scope of the present invention.

[0027] Embodiment 1

[0028] An auxiliary diagnosis system for diabetic retinopathy, where the diabetic retinopathy includes hemorrhage spots, hard exudates, soft exudates, microaneurysms, and optic discs. The auxiliary diagnosis system for diabetic retinopathy includes

[0029] A data acquisition and upload module, which is used to acquire fundus images of patients and upload them to the diabetic retinopathy grading module and the diabetic retinopathy segmentation module;

[0030] A diabetic retinopathy grading module, which is used to grade the received fundus images for diabetic retinopathy, obtain the diabetic retinopathy grade of the fundus images, and transmit it to the intelligent auxiliary diagnosis and analysis module;

[0031] The diabetic retinopathy segmentation module is used to input the received fundus image into the diabetic retinopathy segmentation module for lesion segmentation, obtain the fundus image lesion segmentation map, and transmit it to the intelligent auxiliary diagnosis and analysis module;

[0032] The intelligent auxiliary diagnosis and analysis module is used to comprehensively analyze the received fundus image lesion segmentation map, and generate a graphic report in combination with the received diabetic retinopathy grade of the fundus image, and display it to the user. The comprehensive analysis includes analyzing the type, location, and contour information of diabetic retinopathy.

[0033] The specific diagnosis method of the diabetic retinopathy auxiliary diagnosis system includes the following steps:

[0034] S1. The data acquisition and upload module acquires the original fundus image and performs preprocessing to obtain the preprocessed fundus image, and then uploads the preprocessed fundus image to the diabetic retinopathy grading module and the diabetic retinopathy segmentation module;

[0035] The preprocessing method is: perform data augmentation on the original fundus image to alleviate the problem of low accuracy caused by data imbalance, and can also enhance the generalization ability of the model; data augmentation is performed by using existing methods such as adding data, generating images by GAN, and random rotation for data expansion.

[0036] S2. After receiving the preprocessed fundus image, the diabetic retinopathy grading module uses the pre-trained diabetic retinopathy grading model to grade the preprocessed fundus image for diabetic retinopathy, obtains the diabetic retinopathy grade of the fundus image, and transmits it to the intelligent auxiliary diagnosis and analysis module;

[0037] S3. After receiving the preprocessed fundus image, the diabetic retinopathy segmentation module uses the pre-trained diabetic retinopathy segmentation model to perform image segmentation on the preprocessed fundus image, obtains the fundus image lesion segmentation map, and transmits it to the intelligent auxiliary diagnosis and analysis module;

[0038] S4. After receiving the fundus image lesion segmentation map, the intelligent auxiliary diagnosis and analysis module performs comprehensive analysis to obtain the contour and location information of the lesions in the fundus image, and then generates a graphic report in combination with the received diabetic retinopathy grade of the fundus image, and displays it to the user.

[0039] The original fundus image described in step S1 is obtained by photographing the patient's eye with a fundus camera, and can be a real-time photographed image, or an image stored in a local server or cloud, and is obtained by reading or network transmission.

[0040] The diabetic retinopathy grading model described in step S2 uses the GhostNet network model and grades the preprocessed fundus images according to the grading standard. When training the GhostNet network model, the existing graded fundus image dataset is divided into a training set, a validation set, and a test set. Then, the training set is input into the GhostNet network model for training, the validation set is used for auxiliary verification training, and the test set is used for detection to obtain a trained diabetic retinopathy grading model with a detection accuracy reaching the set requirements.

[0041] The GhostNet classification model has a good effect on classifying fundus images and is a lightweight model with high operating efficiency. The fundus image training set is input into the network for training, the validation set is used for auxiliary training, and finally the test set is input into the trained model for detection to obtain the detection accuracy of the model.

[0042] The GhostNet network model consists of 1 Conv3, 16 G-bneck modules, 1 Conv1, 1 Pooling, and 1 FC. The Conv3 represents a convolutional layer with a convolution kernel size of 3×3 and a stride of 2; the Conv1 represents a convolutional layer with a convolution kernel size of 1×1; the Pooling is an average pooling layer; the FC is a fully connected layer; its process is expressed by the formula

[0043]

[0044] In the formula, F O represents the output feature map; F I represents the input feature map; GM represents the Ghost structure; BN represents batch normalization processing; RL represents the ReLU linear correction unit.

[0045] The G-bneck module is divided into the G-bneck1 module and the G-bneck2 module. The G-bneck1 module consists of 2 Ghost structures, 2 batch normalization layers, and 1 ReLU activation function. The G-bneck2 module consists of 2 Ghost structures, a Depthwise convolutional layer with a stride of 2, 3 batch normalization layers, and 2 ReLU activation functions. Its process is expressed by the formula

[0046]

[0047] In the formula, DWConv represents a Depthwise convolutional layer with a stride of 2.

[0048] Among the 16 G-bneck modules, the 1st, 3rd, 5th, 7th, 8th, 9th, 10th, 11th, 13th, 14th, 15th, and 16th G-bneck modules adopt the G-bneck1 module, and the 2nd, 4th, 6th, and 12th G-bneck modules adopt the G-bneck2 module.

[0049] The method for training GhostNet includes a training method and a loss function. The training method uses the SGD optimizer, and the loss function is the cross-entropy loss function, which is expressed by the formula

[0050] Cross-entropy loss In the formula, p(x) represents the expected output value, q(x) represents the actual output value, n represents the number of samples, and x i represents the i-th input feature sample.

[0051] The diabetic retinopathy segmentation model adopts the DRSU-Net (Diabetes retinopathy segmentation U-Net) network model. When training the DRSU-Net network model, the existing dataset of segmented fundus image lesion segmentation maps is divided into a training set, a validation set, and a test set. Then, the training set is input into the DRSU-Net network model for training, the validation set is used for auxiliary verification training, and the test set is used for detection to obtain a trained diabetic retinopathy segmentation model with a detection accuracy reaching the set requirements. During training, the Adam optimizer is adopted, and the loss function adopts the binary cross-entropy loss function.

[0052] The DRSU-Net network model includes a convolutional layer, a pooling layer, a deconvolutional layer, and an activation function. The pooling layer adopts the max-pooling operation. The activation function adopts the ReLU linear rectification unit. The convolutional layer is used to extract high-dimensional feature information of the image, and the deconvolutional layer is used to restore the feature map to the size of the original input image.

[0053] Assume that the input image size is i×i. Then the output image sizes o c and o dc of the convolutional layer and the deconvolutional layer are respectively

[0054]

[0055] o dc = s*(i - 1) + k - 2*p

[0056] In the formula, o c represents the output image size of the convolutional layer; o dc represents the output image size of the deconvolutional layer; i represents the input image size; k represents the convolutional kernel size; p represents the padding size; s represents the convolutional kernel moving step size.

[0057] The DRSU-Net network model includes an encoder part, a decoder part, a bridging part, and a skip connection part. The encoder part is used to obtain the high-level abstract features of the image; the decoder part is used to restore the feature map to its original size; the bridging part is used to connect the encoder part and the decoder part; the skip connection part is used to fuse the feature information of the encoder part and the decoder part.

[0058] In one embodiment, the encoder module is composed of four downsampling residual modules. The residual structure can improve the performance of the network, alleviate the problem of gradient descent, and retain the shallow information in the deep structure, alleviating the problem of information loss caused by convolution and downsampling. The downsampling residual module consists of a combination of two groups of batch normalization, ReLU activation function, and convolutional layers. Its process is expressed by the formula

[0059]

[0060] In the formula, F OUT represents the output feature map; F IN represents the input feature map; BN represents batch normalization processing; RL represents the ReLU linear correction unit; Conv represents a 3×3 convolutional kernel.

[0061] In one embodiment, the decoder module corresponds to the encoder module and is composed of 4 upsampling residual modules. The upsampling residual module consists of a combination of two groups of batch normalization, ReLU activation function, and deconvolutional layers. Its process is expressed by the formula

[0062]

[0063] In the formula, F A represents the output feature map of the skip connection part; DConv represents a 2×2 deconvolutional kernel.

[0064] In one embodiment, the skip connection part includes an attention mechanism. The attention mechanism can effectively suppress noise and irrelevant information and improve the segmentation performance of the network. The attention mechanism includes a spatial attention module and a channel attention module. The specific process is as follows: First, the feature map is first subjected to global pooling operation, then convolutional operations are performed using two 1×1 convolutional kernels, and then it is sent into the Sigmoid activation function. Then, the input feature map is fused with the result of the Sigmoid function as the output of the channel attention part. Then, the output feature map of the channel attention part is sent into the Sigmoid activation function after pooling operation in the channel dimension. Finally, the residual idea is used to fuse the output feature map of the channel attention part with the result of the Sigmoid function as the output of the entire attention mechanism module. The whole process is expressed by the formula

[0065]

[0066]

[0067] In the formula, H represents the height of the feature map, W represents the width of the feature map, C represents the number of channels of the feature map, σ represents the Sigmoid function, AP represents average pooling, Conv2 represents two 1×1 convolutional kernels, and F CA-IN represents the input feature map of the channel attention mechanism, and F CA-OUT represents the output feature map of the channel attention part, and F SA-IN represents the input feature map of the spatial attention mechanism, C_GAP represents channel average pooling, and F SA-OUT represents the output feature map of the spatial attention mechanism, which is also the output of the attention mechanism module.

[0068] In one embodiment, the bridging part is also composed of a combination of two groups of batch normalization, ReLU activation function, and convolutional layer, and its process is represented by the formula

[0069] F OUT = DConv(Conv(RL(BN(Conv(RL(BN(F IN )))))))

[0070] In some embodiments, the DRSU-Net method includes a training method and a loss function. The training method uses the Adam optimizer, and the loss function is the binary cross-entropy loss function, which is represented by the formula

[0071]

[0072] In the formula, N represents the total number of samples, y i represents the category to which the i-th sample belongs, and p i represents the predicted value of the i-th sample.

[0073] In some embodiments, the qualitative and quantitative analysis includes the analysis of information such as the type, location, and contour of diabetic retinopathy.

[0074] In some embodiments, the diabetic retinopathy grades are divided into five levels according to the international diabetic retinopathy grading standard, including no abnormality, mild non-proliferative diabetic retinopathy, moderate non-proliferative diabetic retinopathy, severe non-proliferative diabetic retinopathy, and proliferative diabetic retinopathy.

[0075] In some embodiments, the graphic report includes information on the diabetic retinopathy grade of the fundus image, the location and contour information of diabetic retinopathy in the fundus image, the diagnosis result, treatment suggestions, etc.

[0076] Example 2

[0077] A diabetic retinopathy assisted diagnosis system includes a digital processing device, which includes: a processor, a memory, and a computer program. The computer program is executed to create an application that uses a deep neural network to perform real-time diagnosis on medical images and give a graphic diagnosis report.

[0078] A data acquisition and upload module for acquiring fundus images of a patient;

[0079] A diabetic retinopathy grading module for grading diabetic retinopathy of the fundus images;

[0080] A diabetic retinopathy segmentation module is built based on a convolutional neural network. The diabetic retinopathy segmentation module is used to input the fundus images into the diabetic retinopathy segmentation module for lesion segmentation;

[0081] An intelligent assisted diagnosis and analysis module automatically analyzes information such as the type, quantity, and size of diabetic retinopathy, and displays it to the user together with the diabetic retinopathy grading result. Finally, a graphic report is generated to assist ophthalmologists in making better diagnoses.

[0082] The data acquisition and upload module is used to acquire fundus image data and store it in the file memory of the server. The acquired fundus images are used for subsequent lesion detection.

[0083] In some embodiments, the fundus images are obtained by photographing the patient's eyes using a fundus camera, and can be real-time photographed images or images stored in a local server or the cloud, and are obtained by reading or network transmission.

[0084] Preferably, the diabetic retinopathy grading module is used to grade the fundus images obtained by the acquisition and upload module using a trained diabetic retinopathy grading network, and obtain the diabetic retinopathy grade of the fundus images.

[0085] In some embodiments, the diabetic retinopathy grading network uses a trained GhostNet classification model.

[0086] In some embodiments, for the trained GhostNet classification model, after the model is built, the fundus image training set is input into the network for training, the validation set is used for auxiliary training, and finally the test set is input into the trained model for detection to obtain the detection accuracy of the model.

[0087] The diabetic retinopathy segmentation module is used to segment the lesions of the fundus images obtained by the acquisition and upload module using the trained diabetic retinopathy grading network, and obtain the diabetic retinopathy segmentation map of the fundus images.

[0088] In some embodiments, for the trained DRSU-Net model, after the model is built, the fundus image training set is input into the network for training, the validation set is used for auxiliary training, and finally the test set is input into the trained model for detection to obtain the detection accuracy of the model.

[0089] Preferably, the intelligent auxiliary diagnosis and analysis module is used to comprehensively analyze the results of the diabetic retinopathy grading module and the diabetic retinopathy segmentation module, automatically generate a diagnosis result, display it to the user on the human-computer interaction interface, and finally generate a graphic report.

[0090] In some embodiments, the diabetic retinopathy includes hemorrhage spots, hard exudates, soft exudates, microaneurysms, optic discs, etc.

[0091] In some embodiments, the qualitative and quantitative analysis includes the analysis of information such as the type, location, and contour of diabetic retinopathy.

[0092] In some embodiments, the diabetic retinopathy grades are divided into five levels according to the international diabetic retinopathy grading standard, including no abnormality, mild non-proliferative diabetic retinopathy, moderate non-proliferative diabetic retinopathy, severe non-proliferative diabetic retinopathy, and proliferative diabetic retinopathy. In some embodiments, the human-computer interaction interface can display information such as the original fundus image, the change grade of the fundus image, the marked map of diabetic retinopathy, and diagnostic suggestions, and can perform operations such as rotation and scaling according to the needs of the user.

[0093] In some embodiments, the system runs in real time and automatically generates a diagnosis result in real time.

[0094] In some embodiments, the graphic report includes all the information in the human-computer interaction interface and the diagnosis results and suggestions of ophthalmologists.

[0095] In some embodiments, the diagnosis results and suggestions of the ophthalmologist are manually filled in by the ophthalmologist after reviewing the human-computer interaction interface.

[0096] As described above, for the method and system for auxiliary diagnosis of diabetic retinopathy of the present application, first, a fundus image is obtained, and then the fundus image is input into a diabetic retinopathy grading model to obtain a lesion grade. Then, the fundus image is input into a diabetic retinopathy segmentation model to obtain a lesion segmentation map. Finally, the above results are comprehensively analyzed to obtain a final diagnosis result, which is beneficial to improving the accuracy, efficiency and intelligent level of the diagnosis of diabetic retinopathy.

[0097] The above embodiments are only exemplary embodiments of the present invention. Within the understanding of those skilled in the art, various changes and modifications within the spirit and principle of the present invention are within the protection scope of the present invention.

Claims

1. An auxiliary diagnosis system for diabetic retinopathy, characterized in that: including a data acquisition and upload module, configured to acquire fundus images of patients and upload them to the diabetic retinopathy grading module and the diabetic retinopathy segmentation module; a diabetic retinopathy grading module, configured to grade the received fundus images for diabetic retinopathy, obtain the diabetic retinopathy grades of the fundus images, and transmit them to the intelligent auxiliary diagnosis and analysis module; the diabetic retinopathy grading model adopts a GhostNet network model and grades the preprocessed fundus images according to the grading standard; when training the GhostNet network model, the existing dataset of fundus images with known grades is divided into a training set, a validation set, and a test set, then the training set is input into the GhostNet network model for training, the validation set is used for auxiliary verification training, and the test set is used for detection to obtain a trained diabetic retinopathy grading model with a detection accuracy reaching the set requirements; the GhostNet network model consists of 1 Conv3, 16 G-bneck modules, 1 Conv1, 1 Pooling, and 1 FC. The Conv3 represents a convolutional layer with a kernel size of 3×3 and a stride of 2; the Conv1 represents a convolutional layer with a kernel size of 1×1; the Pooling is an average pooling layer; the FC is a fully connected layer; the G-bneck module is divided into a G-bneck1 module and a G-bneck2 module. The G-bneck1 module consists of 2 Ghost structures, 2 batch normalization layers, and 1 ReLU activation function. The G-bneck2 module consists of 2 Ghost structures, a Depthwise convolutional layer with a stride of 2, 3 batch normalization layers, and 2 ReLU activation functions. During training, the SGD optimizer is adopted, and the cross-entropy loss function is used as the loss function; a diabetic retinopathy segmentation module, configured to input the received fundus images into the diabetic retinopathy segmentation module for lesion segmentation, obtain the lesion segmentation map of the fundus images, and transmit it to the intelligent auxiliary diagnosis and analysis module; the diabetic retinopathy segmentation model adopts a DRSU-Net network model; when training the DRSU-Net network model, the existing dataset of segmented lesion segmentation maps of fundus images is divided into a training set, a validation set, and a test set, then the training set is input into the DRSU-Nett network model for training, the validation set is used for auxiliary verification training, and the test set is used for detection to obtain a trained diabetic retinopathy segmentation model with a detection accuracy reaching the set requirements. During training, the Adam optimizer is adopted, and the binary cross-entropy loss function is used as the loss function; the DRSU-Net network model includes an encoder part, a decoder part, a bridging part, and a skip connection part. The encoder part is configured to obtain high-level abstract features of the image; the decoder part is configured to restore the feature map to its original size; the bridging part is configured to connect the encoder part and the decoder part; the skip connection part is configured to fuse the feature information of the encoder part and the decoder part; The intelligent auxiliary diagnosis and analysis module is used to comprehensively analyze the received fundus image lesion segmentation map, and generate a graphic report in combination with the received diabetic retinopathy grade of the fundus image, and display it to the user.

2. The auxiliary diagnosis system for diabetic retinopathy according to claim 1, wherein: The diabetic retinopathy mentioned above includes hemorrhage points, hard exudates, soft exudates, microaneurysms, and optic discs.

3. The auxiliary diagnosis system for diabetic retinopathy according to claim 1, wherein: The comprehensive analysis mentioned above includes analyzing the types, locations, and contour information of diabetic retinopathy.

4. The auxiliary diagnosis system for diabetic retinopathy according to claim 1, characterized in that: The specific diagnosis method of the system includes the following steps: S1. The data acquisition and upload module acquires the original fundus image and performs preprocessing to obtain the preprocessed fundus image, and then uploads the preprocessed fundus image to the diabetic retinopathy grading module and the diabetic retinopathy segmentation module; S2. After receiving the preprocessed fundus image, the diabetic retinopathy grading module uses the pre-trained diabetic retinopathy grading model to grade the preprocessed fundus image for diabetic retinopathy, obtains the diabetic retinopathy grade of the fundus image, and transmits it to the intelligent auxiliary diagnosis and analysis module; S3. After receiving the preprocessed fundus image, the diabetic retinopathy segmentation module uses the pre-trained diabetic retinopathy segmentation model to perform image segmentation on the preprocessed fundus image, obtains the fundus image lesion segmentation map, and transmits it to the intelligent auxiliary diagnosis and analysis module; S4. After receiving the fundus image lesion segmentation map, the intelligent auxiliary diagnosis and analysis module performs comprehensive analysis to obtain the contour and location information of the lesions in the fundus image, and then generates a graphic report in combination with the received diabetic retinopathy grade of the fundus image, and displays it to the user.

5. An auxiliary diagnosis system for diabetic retinopathy according to claim 4, wherein: The original fundus image in step S1 is obtained by photographing the patient's eye with a fundus camera, which is a real-time photographed image or an image stored in a local server or the cloud, and is obtained by reading or network transmission.

6. The auxiliary diagnosis system for diabetic retinopathy according to claim 1, wherein: The DRSU-Net network model includes a convolutional layer, a pooling layer, a deconvolutional layer, and an activation function; the pooling layer uses a max pooling operation; the activation function uses a ReLU linear rectification unit; the convolutional layer is used to extract high-dimensional feature information of the image, and the deconvolutional layer is used to restore the feature map to the size of the original input image.

Images