Diabetic Retinopathy Grade Screening System Based On Deep Learning

Diabetic Retinopathy(DR)is a susceptible eye disease in diabetics that requires early diagnosis and treatment to avoid affecting the patient’s vision.There are many ways to diagnose DR,and fundus photography is more intuitive and simple than other examinations,and fundus camera equipment is relatively inexpensive,which is the most suitable method for diabetic retinopathy.At this stage,the diagnosis of diabetic retinopathy basically relies on the naked eye observation of fundus images by ophthalmologists,and the accuracy of examination depends on the medical skills of ophthalmologists.However,the manual diagnosis of doctors is difficult to meet the needs of a large number of diabetic people,and long-term examination is easy to fatigue,affecting the accuracy of disease judgment.For this reason,if computer-aided diagnosis can help doctors complete large-scale DR image analysis and classification,it can not only save a lot of human resources,but also provide the necessary basic conditions for diabetic patients to implement DR census.This article will explore the establishment of a deep learning-based automatic classification network for diabetic retinopathy images.According to the clinical classification of diabetic retinopathy images with different grades,the diabetic retinopathy images taken by the input fundus camera are classified and tested,and the best performance classification network is selected.In this thesis,two models of deep learning,Res Net and Efficient Net V2,are selected based on transfer learning training and the effect of image classification of diabetic retinopathy is tested.In view of the image characteristics of diabetic retinopathy,in order to strengthen the classification performance of the model for the image features of lesions,in addition to studying the classification effect of the residual network without depth and the Efficient Net V2 network with different parameters,this thesis proposes to add attention mechanism to the characteristics of DR images and improve the feature extraction ability to verify the optimization effect of attention mechanism on different networks.Finally,the feasibility of the obtained model in clinically assisting in the diagnosis of diabetic retinopathy is explored.Through experimental results,it is concluded that in the image classification of diabetic retinopathy,Res Net and Efficient Net V2 can have good performance,and the addition of attention mechanism has obvious effect on enhancing image feature extraction ability and improving the performance of model classification.In this thesis,the image classification accuracy of the residual model with the addition channel attention mechanism reached 90.4%,the accuracy rate reached 90.8%,and the recall rate reached 91.0%.In this thesis,the Efficient Net V2 model with the addition of the spatial attention mechanism module has a classification accuracy rate of 89.3%,an accuracy rate of 89.6%,and a recall rate of 90.2%.The area under the working characteristics curve of the improved Res Net model and the Efficient Net V2 model is 98.1% and 97.7%,respectively.The experiments show that the model designed in this thesis has potential application value in the DR screening of diabetic patients.