The Research And Implementation Of Automatic Diabetic Retinopathy Grading And Pathology Location Detection System Based On Deep Learning

Diabetic retinopathy is one of the high incidence of complications for diabetes mellitus,and it is also the main cause of blindness in diabetic patients.The early phases of diabetic retinopathy will not lead to vision decline,but it will be difficult to be cured at the late phases.Therefore,regular follow-up of diabetic patients can effectively stop the deterioration of eye conditions.Due to the shortage of medical resources and the large number of patients,it is very important to study and design an automatic DR grading and pathology location detection system to help doctors improving the efficiency of diagnosis.In recent years,deep learning has been extensively studied and applied in a variety of medical image classifications.The main research content of this thesis is to apply the deep learning algorithm to the intelligent diagnosis of diabetic retinopathy fundus image.The specific work is as follows:(1)In this thesis,convolutional neural network is used to classify the stage 4 diabetic retinopathy.Improve the Inception V3 model,add the flat layer,dense layer,targeted dropout layer,dense layer and softmax layer in sequence after the mixed layer in the original network structure.This model greatly improves the accuracy of stage 4 diabetic retinopathy in the stage of diabetic retinopathy.(2)Aiming at these problems that the size of diabetic retinopathy images are too large compare with the small focus,the lesion location are too scattered in the image,and some tissues in retina will affect the detection of focus,this thesis proposes a diabetic retinopathy grading method based on improved Faster R-CNN and sub image segmentation technology.Firstly,the target detection network is used to locate the coordinates of the optic disc region,therefore,the influence of the optic disc area on the recognition of lesions is solved;Secondly,the sub image segmentation technology is used for the input image and the depth residual network is used in the feature extraction,then the problem that the focus part is too small to affect the feature extraction is solved;Finally,the interesting region of the image is generated by using online difficult sample mining method,which can solve the problem of imbalance between positive and negative hard and easy samples.(3)Aiming at the problem that the amount of artificially labeled diabetic retinopathy data is small,the idea of self-training to extract pseudo truth value is adopted in this thesis,and semi-supervised learning is used to introduce false labels for iterative training during the network training stage.Based on the trained initial network model for tuning training,the parameters and model of the algorithm were finally determined.The diabetic retinopathy staging experiment was performed on the Eye PACS data set and diabetic retinopathy fundus samples from Aier Eye Hospital.The accuracy of the proposed method in diabetic retinopathy staging reached 95.52% in Phase 0,86.35% in Phase 1,and 93.52 in Phase 2,87.50% in Phase 3,83.91% in Phase 4.(4)Finally,an algorithm for diabetic retinopathy fundus image was established,the demand analysis of the intelligent diagnosis system for diabetic retinopathy was completed,the system is designed and realized.