Detection Of Diabetic Retinopathy Based On Attention Mechanism

Diabetic retinopathy(DR)is one of the complications of diabetes.Although ophthalmologist can analyze fundus images of diabetic patients to determine the disease condition,there are a large number of diabetic patients in China and a relative shortage of ophthalmic experts,which will lead to patients easy to miss the best treatment period,and eventually lead to permanent loss of vision.Relying on computer vision assisted diagnosis technology can reduce the initial diagnosis period of patients and significantly improve the efficiency of DR detection.Therefore,the research on DR automatic detection method has very important practical significance.In this dissertation,convolutional neural network(CNN)is combined with different attention models to study the DR automatic detection method,so as to achieve the goal of DR classification and hard exudate detection.Aiming at the problems of difficult extraction of lesion features and low grading Accuracy in diabetic retinopathy,this dissertation proposed a block-channel domain attention network model(CCAB-Net)to grade fundus images of diabetic patients.The network model is composed of Inception-ResNet-V2 backbone network,improved HSAM attention enhancement module and new CCAB classification attention module.The backbone network is used for feature extraction,the HSAM module is used to enhance the useful information in the feature and suppress the useless information,and the CCAB module is used to block the feature from the channel domain to expand the distance between different categories and improve the Accuracy of image classification.In this dissertation,five-classification experiments were conducted on Eye PACS dataset and five-classification and binary classification experiments on APTOS-2019 dataset.The results show that CCAB-Net improves the evaluation indexes of DR classification and can classify DR images well.Hard exudate is the most representative lesion feature of DR.accurate segmentation of hard exudate plays an important role in the diagnosis and treatment of DR.In this dissertation,a two-stream learning model TS-MAF integrating multi-scale attention modules is designed to segment hard exudates in DR images on the basis of binary classification in DR classification tasks.The model consists of two streams of image segmentation stream and super resolution stream,and uses multi-scale attention fusion module to integrate information,which can effectively improve the Accuracy of small lesions and boundary detection.The results of experimental analysis on IDRi D and E-Ophtha data sets show that the TS-MAF model designed in this dissertation has better segmentation effect on hard exudates,which proves the validity of the TS-MAF model.