Research On Automatic Diagnosis Method Of Diabetic Retinopathy Based On FP Image

Diabetic Retinopathy(DR)is a kind of fundus disease caused by diabetes.The disease is characterized by no symptoms in the early stage,but with the extension of the course of diabetes it will gradually become worse,and even lead to direct blindness.Therefore,it is of great significance for the treatment of this disease for the patient to seek medical advice in time.Hard exudate is a lesion in the fundus of the eye in the early stages of diabetic retinal disease.In fact,clinicians diagnose diabetic retinopathy by observing the status of related lesions in fundus images.Therefore,accurate segmentation of the hard exudate in the fundus images is crucial to the diagnosis of DR disease.Computer-aided diagnosis is a process of developing an automatic diagnosis system based on deep learning to assist doctors in diagnosis.Compared with the time-consuming and laborious difficulty of manual diagnosis by doctors,it is of great practical significance for doctors and patients to add computer aided diagnosis system to realize automatic diagnosis and analysis of retinal fundus images for diabetic retinal diseases.Therefore,based on deep learning technology,this paper developed a system for automatic diagnosis of diabetic retinopathy in fundus images obtained by fundus photography.The main research contents include semantic segmentation of hard exudate and automatic classification of diabetic retinal diseases.Firstly,the semantic segmentation method of hard exudate in fundus images was studied,and a series of research methods were proposed according to the segmentation difficulties of hard exudate,such as scattered location,small target,irregular shape and unbalanced category.Including targeted preprocessing methods and semantic segmentation network model based on U-NET codec network.The model is improved and optimized by introducing cyclic residual module,attention mechanism and deep supervision to improve the segmentation accuracy.Secondly,a network model is proposed to automatically classify the severity of diabetic retinopathy.This model adopts the basic architecture of Res Net and adds Inception structure to extract the multi-scale information of the image,and the final classification result is obtained by integrating the results of the auxiliary classifier.Experimental results show that the proposed method can achieve classification accurately and improve the accuracy and AUC value of classification.