Detection Of Diabetic Retinopathy Based On Deep Learning Method

Diabetes retinopathy(DR)is a fundus disease caused by diabetes.Although the disease has no obvious symptoms at the initial stage,it will become more and more serious as the time of disease increases.According to statistics,the DR incidence rate accounts for about 50% of diabetes patients,and the blindness rate is 6-8%.The visual impairment caused by DR is irreversible,so regular screening for diabetic retinopathy and early detection and treatment are critical.The traditional diagnosis of fundus lesions relies on doctors’ clinical experience and is time-consuming and laborious,making it unsuitable for large-scale screening.Using computer-aided diagnostic systems can not only save doctors’ time,but also discover more potential patients,which has very important practical significance for both doctors and patients.This paper designs two hierarchical networks based on deep learning technology to assist doctors in diagnosis.The main work of this paper is as follows:First,in order to solve the problem that a single normalization method cannot effectively exert the network performance when extracting different layers of features,this paper introduces an adaptive normalization(SN)method in convolutional neural networks to replace the original normalization method.This method can provide different normalization methods when the neural network extracts different layers of features,widen the technical boundary of the network,and make better use of the network performance.In order to better evaluate the authenticity of the model,the Gradient Weighted Class Activation Mapping(Grad-CAM)method is used to generate a prediction thermodynamic map,visually displaying the prediction results of the convolutional neural network,and verifying the authenticity of the network prediction.The obtained thermal location map can also help us better analyze the focus of the network when extracting features,and point out the direction for the next multi lesion optimization.After using the SN method,the network accuracy rate is 88.06%,the sensitivity is 73.46%,and the specificity is 96.88%.All indicators are improved compared to the original network.Secondly,on the basis of summarizing the previous experiment,we designed a grading network of diabetes retinopathy with better performance.To solve the problem of small difference in image features and relatively fuzzy classification threshold between diabetes retinopathy categories.In this chapter,the Conv Ne Xt automatic grading model of diabetes retinopathy is designed,which integrates the attention characteristics of efficient channels.To solve the problem of uneven distribution of samples in various categories of data sets,additional data sets and upsampling methods are introduced to balance the sample distribution.Use preprocessing methods to improve image quality for exposure,artifacts,and other issues that occur when capturing fundus images.Use Dropout and Drop Path methods to alleviate network overfitting problems.Accelerate model training using transfer learning methods.The experimental results show that the accuracy of the model proposed in this paper is95.21%,the sensitivity is 95.20%,and the specificity is 98.80%.The improved convolution neural network can perform the grading task of diabetic retinopathy better.