Research On Diabetic Retinopathy Using Deep Learning Method

In recent years,diabetes has become one of the diseases that seriously affect people’s health.According to statistics,by the end of 2016,the number of people suffering from diabetes in China has reached 110 million,accounting for 10% of the domestic adult population.Diabetes is a chronic disease with a high prevalence and is often accompanied by many serious complications such as diabetic cardiovascular complications and diabetic retinopathy.However,early diabetes is not easy to detect,but exudation,hemorrhage and microaneurysm,which are signs of diabetic retinopathy,can be found early.So the detection of these lesions is now widely used in diabetes screening by the World Health Organization(WHO).So it’s very meaningful to design a computer program to help the diabetic retinopathy srceening as our country has less doctors.As deep learning method has made a great contribution in image recognition,this article uses deep learning method to analyse diabetic retinopathy.Exudation is one of main diabetic retinopahty.Exudation detection is very useful for diabetes screening and clinical diagnosis.Using deep learning method,this article designs an Unet-based convolutional neural network with multi output to detect the exudation location.In view of the imbalance of positive and negative samples,we introduce the L1-based hard samples mining to make the training process faster.The AUC of the segmentation network under the ROC curves measured on three public databases DIARETDB1,HEI-MED,and Messidor are0.96,0.94 and 0.91,respectively.Compared with other work,the method in this article achieves the best results on the DIARETDB1 database.In the process of diabetes screening,doctors usually check whether the fundus images have lesions or grade the degree of diabetic retinopathy.So a big data based deep learning method is proposed in this article.This paper designs a 25-layer convolution neural network to extract the lesion features and a 5-layer multi-layer perception as the final lesion classifier.The idea of iterative training is proposed for training the deep neural networks,which means training network parameters from shallow to deep in turn by three steps.We use image enhancement to transform one image into multiple images and then extract the mean and variance of multiple images’ features as input to the classifier.In addition,we also use the method of model fusion,that is,designing two classification networks,and averaging their results as the final output of the system.The classification system achieve an accuracy of 55.13% in a subset of 6000 images in the Kaggle database.At the same time,the binary(diseased and normal)test has been done on the public dataset e_optha,yielding a sensitivity of 94.36% and a specificity of 88%.