Research On Image Recognition Of Diabetic Retinopathy Based On Deep Learning

Diabetic Retinopathy(DR)is a retinal disease caused by diabetes.There are many people worldwide who have become blind due to DR,so early detection and treatment of DR is necessary to delay or prevent vision deterioration and loss.Currently,deep learning-based fundus image analysis and eye disease detection technology has been widely studied and focused on by scholars.Doctors diagnose DR by observing the curvature,length,width,branching and other features of the retinal blood vessels;grading DR can help patients better understand their condition in order to better comply with doctors’ treatment recommendations.However,retinal vessels image segmentation and DR classification are demanding and timeconsuming for medical personnel.Therefore,the use of deep learning technology to achieve automatic segmentation of retinal vessels image and classification of DR is of great significance for clinical diagnosis.The main research of this paper is as follows:1.Aiming at the problem that the traditional retinal vascular image segmentation network loses tiny feature information as the network depth deepens,and the network segmentation sensitivity is low,a residual channel attention(Residual Channel Attention,RCA)module and multi-scale dilated convolution(Multi-scale Dilated Convolution,MDC)module is used as the basic feature extraction module network for fundus blood vessel image segmentation.The maximum number of channel layers of the feature map is only 64,and the feature map size halving and deconvolution operations are only twice,which can reduce the information loss phenomenon caused by the change of the feature map size.The accuracy of the proposed method tested on DRIVE and CHASE-DB1 datasets is 96.85% and 97.39%,the sensitivity is84.03% and 86.50%,respectively.2.In order to meet the storage capacity of the handheld fundus camera and the requirement of increasing the screening speed in the DR screening process,the RCA-MDC lightweight network was used as the feature extractor for DR binary classification,doubling the number of channels per layer and changing its output structure,feeding the pre-processed fundus images into the network for feature extraction,and then feeding the extracted features into the Multi-Layer perceptron(MLP)designed in this paper.The experimental results showed that the accuracy of the network reached 95.82%,and the weight parameter of the whole network was only 17.3 MB,which was less than 1.5% compared with the accuracy obtained by using VGG and other classical networks using transfer learning for binary classification,but the weight parameter was less than one-tenth of the weight of these classical networks.Considering that the doctor can determine whether the patient has diabetic retinopathy by the shape of blood vessels in the fundus image,this paper also feeds the presegmented fundus image into the RCA-MDC lightweight network for feature extraction,and then fuses the features obtained from both inputs into the designed Multi-Layer Perceptron(MLP)for DR biclassification.The experimental results show that the network with twochannel input achieves 96.36% accuracy in biclassification.3.The integration of DR five classification on a desktop fundus camera is not limited by the processing speed and the cost of hardware.Since lightweight networks cannot meet the requirements of DR five classification applications,in order to implement DR five classification to provide more comprehensive and personalized support for diabetes management,thus reducing the risk of lesions and the cost of diabetes treatment.The parameters of the trained deep neural network model on Image Net dataset are transferred to our own neural network model using the related techniques of transfer learning,and then feature extraction and classification are performed on this basis.A DR five-classification network based on Efficient Net B7,Xception network and channel attention module as feature extractors is designed to solve the five-classification problem of DR lesions.The experimental results show that the DR five-classification accuracy obtained by this network on the APTOS test set reaches 84.73%.In this paper,A miniaturized RCA-MDC network was designed to achieve accurate segmentation of retinal vessels image,and used the RCA-MDC network to achieve DR biclassification to discriminate whether fundus images are suffering from diseases.In order to realize the function of DR five classification,the parameters of the trained deep neural network model on the Image Net dataset are migrated to their own neural network model using the related technique of Transfer learning,and then feature extraction and classification are performed on this basis.A DR five-classification network based on Efficient Net B7,Xception network and channel attention module as feature extractors is designed to solve the DR lesion five-classification problem.