A Research And Implementation Of Classification Of Diabetic Retinopathy Of Fundus Images

In clinical medicine,retinal fundus image is the main basis for ophthalmologists to diagnose and treat patients with diabetic retinopathy.This diagnosis process is timeconsuming and labor-intensive and its accuracy depends heavily on the professionalism of the doctor,so it is unable to meet the rising patient demand.As a result,many patients missed the best treatment opportunity.Therefore,the research and implementation of automatic classification of diabetic retinopathy based on fundus images is of great significance for the timely detection and treatment of diabetic retinopathy.This thesis is mainly about the design and implementation of a classification model for diabetic retinopathy that solves the problem of the unbalanced data set of retinal fundus images,sparsely distributed lesion features in retinal fundus images,and the lack of explanatory of deep learning models.Based on deep learning and active learning,an accurate and efficient classification model of diabetic retinopathy is studied and designed.The main research contents of this thesis include the following three aspects:1.Aiming at the the unbalanced dataset of fundus images,A improved AIFT(active,incremental fine-tuning)based on multi-task network model is proposed.The method of active samples selection,handling noisy samples and continuous fine-tuning of AIFT are improved.Experiments show that compared with the AIFT algorithm,the multi-task network model with the improved AIFT algorithm can save 20% of the training time,and can reduce the training time by 40% compared with the random selection of samples for incremental training.2.Aiming at the problem that the distribution of lesion features in the fundus image is sparse,and the traditional deep learning model has low interpretability,a diabetic retinopathy classification network based on lesion detection is proposed.First,a lesion detection network was designed to detect key lesions(bleeding,exudation and microaneurysms)in the fundus image,which improved the interpretability of the model.The diabetic retinopathy classification model uses the results of lesion detection to correct the result of the multi-task network,and the experiments show that the accuracy of the model reached 79%,and the sensitivity reached 68%.3.A diabetic retinopathy classification system was designed and implemented,which realized the automatic diagnosis and management of the diabetic retinopathy,and patient’s diagnosis of diabetic retinopathy and lesion detection results are visualized.In the actual clinical test,the system reflects a high value of use.