Deep Learning-based Research On The Recognition Technology For Retinal Lesions In Diabetes

Diabetic retinopathy is a serious and highly prevalent syndrome of diabetes mellitus,manifested as microangiopathy of the eye,which can affect vision and even blindness.Currently,the diagnosis of this condition relies on the analysis of fundus images by ophthalmologists to determine the type and location of the lesion and to give diagnostic and treatment advice.However,at this stage,the supply and demand of doctors and patients in China are seriously imbalanced,and patients in remote areas cannot receive timely diagnosis and treatment due to lack of health resources.Therefore,the design of an auxiliary diagnostic system that can automatically identify and segment diabetic retinal lesions is of great importance for clinical medical research and patient vision protection.In this thesis,we investigate the technical variation of diabetic retinal lesion identification,and the main work is as follows.(1)To address the problem of uneven scale distribution of diabetic retinal lesions in fundus images and complex background tissue information in the fundus,which affects target lesion feature extraction,a segmentation model based on multi-channel residual convolution module and gated attention module is proposed on the basis of U-Net.The model is designed with a multi-channel residual convolution module for fusing global semantic information of different sensory regions and local detail information of different sizes to enhance the network’s ability to capture multi-scale target features;in addition,the model uses a gated attention module to suppress feature responses from regions unrelated to or less relevant to the target,enhancing the network’s ability to recognize and distinguish background regions.The experimental results show that the model can accurately segment bleeds and exudates,and the experimental effect is greatly improved.(2)A semi-supervised segmentation model based on Mean Teacher combined with an uncertainty-aware strategy is proposed to address the problem that the number of labeled samples of fundus images is sparse and difficult to meet the training of deep network models.Based on the previous study,the semi-supervised segmentation model Mean Teacher is introduced,which is able to achieve better segmentation performance in the absence of labeled samples by learning the commonality among unlabeled data.In addition,to further improve the prediction accuracy of the model for unlabeled data,an uncertainty-aware strategy is introduced to guide the model to focus on the uncertainty of the prediction region and obtain more reliable prediction results.The experimental results show that the model can effectively improve the recognition of diabetic retinal lesions.(3)Based on the previous study,a diabetic retinal lesion assisted recognition system was designed and implemented.This system can complete the localization and segmentation of diabetic retinal lesions,help patients and medical personnel to obtain various types of lesion information in fundus images,and provide decision basis for doctors’ treatment plan formulation.