Research And Application Of Convolutional Neural Network In The Classification Of Diabetic Retinopathy

At present,the high incidence of diabetes has become a social disease,Diabetic retinopathy(DR)is a common complication of diabetes melitus,often leads to the patient's blindness.The process of diagnosing diabetic retinopathy is time-consuming and laborious,leading to the failure of timely diagnosis in most patients.In recent years,with the development of deep learning,it is possible to apply it in the auxiliary diagnosis of diabetic retinopathy.This paper takes diabetic retinopathy as the research object and establishes an algorithm model for the classification of diabetic retinopathy.Firstly,according to the noise of fundus image,a series of preprocessing and data enhancement methods are proposed,and the loss function weighting method is adopted to deal with the problem of unbalanced data set.Then,respectively,using the Inception structure feature extraction ability,increase the convolution network width DenseNet reuse characteristics in deep convolution network ability,and lightweight convolution MobileNet network advantages,retraining of above three convolutional neural network model,to classify the severity of diabetic retinopathy,and use the characteristics of different pathological changes of the sensitivity of different model,integration the model in order to improve the final accuracy.Finally,in view of the difficulties in grassroots screening of diabetic retinopathy and the high rental cost of cloud services,this paper designed a simple auxiliary diagnosis system by taking advantage of the portability and simple operation of mobile terminals,and deployed the model on the system to provide diabetes patients and medical staff with auxiliary diagnosis Suggestions.In this paper,the traditional image processing method is used to preprocess the fundus image to improve the learning efficiency of the model.By integrating multiple models,the accuracy and specificity were improved by 2.91% and 4.37% respectively compared with the optimal model before integration.The algorithm model and image processing method are deployed on the mobile terminal to improve the screening efficiency of diabetic retinopathy and make the research results of practical significance.