Research On Auxiliary Diagnosis Method Of Diabetic Retinopathy Based On Deep Learning

Diabetic retinopathy(DR)is a common complication of diabetes and a worldwide public health problem that seriously threatens human health.Due to the diversity and complexity of diabetic retinopathy,it is difficult to perform DR detection in time-consuming manual diagnosis,so there is an urgent need to establish an automated diabetic retinopathy auxiliary diagnosis model.This topic is to carry out screening and auxiliary diagnosis of diabetic retinopathy through automatic analysis of fundus images under this background.This article uses the deep learning method,using the publicly available Kaggle data set as the diabetic retinopathy data set,through the convolutional neural network to screen diabetic retinopathy,which includes the following main steps: image preprocessing,training residuals of different depths Neural network model,training Google Net inception V3 model,verifying the accuracy of the model,screening and diagnosis of diabetic retinopathy on the resnet50 model,mainly completed the following research content:The image is preprocessed first.In view of the uneven image quality of the data set and the uneven image of the disease course category,the image quality of the fundus is enhanced through methods such as image quality evaluation,filtering,contrast enhancement,sharpening,and data amplification.Then build the residual neural network model.This paper builds residual neural network models of different depths,uses a large number of processed fundus images as the training set training model,and verifies the best model that can effectively diagnose and grade diabetic retinopathy on the validation set.Experiments have proved that in the residual neural network models of different depths,too deep or too shallow can not get the best results.This article has achieved the best results on the resnet50 model.On this model,the maximum accuracy rate for diagnosing diabetic retinopathy is 96.62%;the maximum accuracy rate for grading diabetic retinopathy is 97.25%.Then build the Google Net inception V3 model.The residual neural network mainly improves the performance of the network by deepening the number of network layers and learning the difference between the input and output.Therefore,this chapter selects Googlenet which increases the width of the network to compare with the residual neural network,and compares the neural network models of different structures for diabetes The pros and cons of the diagnosis of retinopathy.The best accuracy rate of Googlenet-Inception V3 model in diagnosing whether diabetic retinopathy is 95.17%,and the best accuracy rate for grading diabetic retinopathy is 95.53%.Through experimental comparison,Inception V3 model has a deeper residual neural network.The accuracy of the model is similar,but it is not as accurate as the resnet50 model.Finally,the new fundus images were screened and diagnosed for diabetic retinopathy on the resnet50 model.Input a single fundus image into the trained resnet50 model,and output the model diagnosis result and the diagnosis result score.Comparing the diagnosis result obtained by the model with the artificially annotated diagnosis result,the accuracy rate of diagnosing diabetic retinopathy is 98%,and the accuracy rate of grading diabetic retinopathy is 95%,which proves that the output result of the model is as good as manual annotation A high degree of consistency.