Classification Of Diabetic Retinopathy Grades Based On Convolutional Neural Network

Diabetic retinopathy is a chronic complication of diabetes that damages the retina.The risk of blindness in diabetic retinopathy patients is 25 times that of healthy people,so diabetic retinopathy is one of the leading causes of blindness in humans.Blindness caused by diabetic retinopathy can be prevented with regular fundus examinations.The analysis shows that at least 90% of people with diabetic retinopathy can minimize their risk of blindness if they are properly treated and prevented.Now,the examination of diabetic retinopathy is very common in developed countries,but the situation in our country is not very optimistic,the ratio of patients to ophthalmologists in China is 1000:1,resulting in very low screening rates for diabetic retinopathy and missed optimal treatment periods.In addition,most patients do not know the serious consequences of diabetic retinopathy,and in many cases this serious complication is ignored.Many patients have reached an advanced stage of treatment,with severe disease,poor treatment effects and high medical costs.Blindness caused by diabetic retinopathy is irreversible,bringing a huge burden to Chinese society and families.Therefore,it is urgent to find new ways to solve this problem.At present,due to the advancement of artificial intelligence,the detection of diabetic retinopathy through automated systems has become a trend in modern medical diagnosis.In this paper,the convolutional neural network is used to classify the grades of diabetic retinopathy with the public color fundus images as the research object.Diabetic retinopathy is divided into five grades: normal,mild,moderate,severe,and proliferative diabetic retinopathy by different degrees of lesions such as microaneurysms,bleeding points,soft and hard exudates,and new blood vessels.The main research contents are as follows:1.Aiming at the difficulty of network classification due to the complexity of diabetic retinopathy information,this paper proposes a convolutional neural network based on dual attention module.First,this paper adopts Resnet50 as the backbone network.Next,in order for the network to effectively capture the features of cross-channel interactions and learn complex lesion information efficiently,we add an ECA attention module in each encoding layer.Finally,the CBAM attention module is added after the last layer Layer4 in the encoding stage,and the channel and spatial attention are used to obtain high-level semantic feature information on the channel and space.According to a large number of comparative experiments and result analysis,the dual attention modules acting on different dimensions in the network can effectively improve the performance of the classification of diabetic retinopathy grades.The accuracy of the proposed DA-Net model for the classification of diabetic retinopathy grades reaches 85.2%,outperforming other models.2.Scholars in the field of medical image processing often use multi-task learning to explore the relationship between different tasks,and improve the performance of each task through information sharing between different tasks.Macular edema is a complication of diabetic retinopathy,and the lesion information of the two diseases overlaps.In order to improve the accuracy of the two diseases.Designing a bilateral crossover network for multitask classification of diabetic retinopathy and macular edema.The BC-Net model uses the backbone network to obtain the feature map,and the feature map distribution is input into two branches,one branch learns the disease characteristics of diabetic retinopathy,and the other branch learns the disease characteristics of macular edema.Then,the obtained specific features of the two diseases are cross-input to the AM aggregation module to learn the intrinsic relationship between the two disease features,and finally the two diseases are classified by a fully connected layer.The experimental data on the Messidor dataset showed that the accuracy rate of the classification of diabetic retinopathy grades reached 91.5%,and the accuracy of the classification of diabetic macular edema grades reached 89.7%.