Recognition Of Diabetic Retinopathy In Images Based On Deep Learning

Color fundus retinal images are the basis for the diagnosis of fundus diseases.Diabetic Retinopathy(DR)is one of the most common types of these fundus diseases.The rapid and accurate processing of fundus images can effectively help diabetic patients in early detection and treatment,thereby greatly improving the cure rate.However,the current manual diagnosis largely relies on experienced doctors,which is not only timeconsuming,but also makes it impossible for patients who lack medical resources in remote areas to receive timely assistance.Therefore,the realization of automated auxiliary diagnosis through image processing,machine learning and deep learning methods is of great significance for the prevention and timely treatment of fundus diseases.This article mainly adopts the method of deep learning to study the following three aspects:(1)Classification of severity of diabetic retinopathy.In diabetic patients,diabetic retinopathy is the leading cause of human blindness.In order to solve the problem that microaneurysms and other tiny lessions in fundus images are difficult to extract,this paper proposes an attention mechanism module.This module adds the weight of the network to the tiny lession by fusing the original feature information of the feature map with the channel information obtained by the attention unit,and then uses the division operation to remove the redundant information in the feature map.Finally the attention mechanism feature as a dual task input;In view of the difficulty in optimization of the Mean Square Error(MSE)loss and the Cross Entropy(CE)loss,the cost of misclassification of DR levels is not considered,a multi-task learning module is designed.The module weighted integrate the MSE loss of the regression task and the CE loss of the classification task.Based on the design of two modules,a fusion of attention mechanism and multi-tasking learning network(FAMT)is proposed.In this paper,after a series of image preprocessing operations such as contrast enhancement and black area removal,the model is verified on a large dataset.(2)Segmentation of lessions of diabetic retinopathy.The severity of diabetic retinopathy depends on several lessions,including microaneurysms(MA),soft exudates(SE),hard exudates(EX)and hemorrhages(HE)etc.the detection of these four types of lesion areas will help provide medical evidence for the classification and diagnosis of diabetic retinopathy.Aiming at the problem of small pixel ratio and multi-size feature capture in fundus images,based on the analysis and verification of mainstream segmentation models,a new hierarchical multi-scale spatial attention gating mechanism(HASG)is proposed.DR segmentation model.In the training phase,the HSAG model inputs multiple multi-scale images to the backbone network for semantic information extraction,and then uses the proposed spatial attention gating unit(SAG)to gradually fuse low-level spatial information and high-level semantic information.The obtained feature maps are distributed with hierarchical attention weights to obtain the final segmentation prediction;in the inference stage,the weights obtained by HASG training are applied in a hierarchical calculation method,and the N prediction scales are combined and calculated to achieve the final segmentation prediction.On the basis of increasing the training cost,the purpose of increasing the segmentation accuracy(3)Diabetic retinopathy identification assist system.Based on the models of the first two studies,a web-based diabetic retinopathy identification auxiliary system was designed and implemented,which is convenient for diabetic patients and medical screening personnel to identify and detect fundus image lesions.This system is based on VUE+Antd as the front-end design technology,combined with Fast Api to deploy the deep learning model on the cloud server,send image segmentation and classification requests through the client,and use the two models deployed in the server to return the results to the user and perform show.