Lesions Segmentation And Disease Course Classification Of Diabetic Retinopathy Based On Convolutional Neural Networks

Diabetic Retinopathy(DR)is a complication of the eye caused by diabetes.It is not easily detectable in the early stages of the disease,but over time it can become more severe and can even lead to blindness.Regular fundus screening in diabetic patients helps patients to detect the disease and provide targeted treatment in a timely manner.In addition,in practice,doctors diagnose the disease by observing the type and number of lesions in the fundus images of patients,so the establishment of an automatic segmentation of DR lesions and classification of DR based on a computer-aided diagnosis system is of great importance to both patients and doctors.The main work of this thesis is as follows:(1)A series of pre-processing methods and a U-Net network-based fundus lesion segmentation model are proposed to address the difficult segmentation problems such as uneven brightness and small sample size of fundus images.The model solves the problem of information loss during the downsampling of fundus lesion segmentation by fusing the attention mechanism and making full use of the information in the feature space domain and channel domain,and effectively extracts the fundus lesion features.The effectiveness of the proposed method is illustrated by the IDRi D dataset.(2)An automatic grading network model for diabetic retinopathy needs referral is proposed,which uses Res Net as the base architecture,incorporates the SE-Block structure to fully extract feature information,and avoids the overfitting problem caused by small samples through the transfer learning method.The proposed method has been validated with the IDRi D dataset and ZDR dataset,and can achieve classification efficiently and accurately.