Research On Fundus Image Analysis And Detection Technology Of Diabetic Retinopathy Based On Deep Learning

At present,retinal fundus image analysis and eye disease detection technology based on deep learning has been studied and paid attention to by a large number of scholars.Retinal blood vessels provide a lot of important information for the diagnosis of diabetic retinopathy(DR)and other ophthalmic diseases.The segmentation result of retinal blood vessels can intuitively show whether the patient’s eyes have lesions,thereby reducing the doctor’s misjudgment of early ophthalmic diseases;And the grading result of DR allows diabetic patients to clearly understand the condition of their fundus lesions,so as to better cooperate with doctors for corresponding treatment.However,retinal blood vessel segmentation and DR grading are not only extremely demanding for medical personnel,but also time-consuming and laborious.Therefore,the use of deep learning technology to achieve automatic retinal vessel segmentation and DR classification under fundus images is of great significance for the clinical diagnosis of DR.The main research work of the paper is as follows:(1)Aiming at the difficulties of the current retinal vessel segmentation task,a twostage training segmentation model based on U-Net network is proposed.Among them,Att Res U-Net network of the first stage is to replace the convolution module of the encoder and decoder in the original U-Net network with a residual module with identity mapping,and attention mechanism is added at every skip-connection;The MiniAtt Res U-net network of the second stage is obtained by tailoring the Att Res U-Net network,which is equivalent to the middle layer of Att Res U-Net.These improvements allow the network to better distinguish between blood vessels and non-vessels,to prevent the loss of small blood vessel pixels,and retain vessel details to a greater extent to improve the segmentation accuracy.The method is tested on two public data sets of DRIVE and STARE,the F1 scores of 0.8351 and 0.8639 and the accuracy rates of 0.969 and 0.978 were obtained respectively.(2)Propose a DR classification method based on SERA-Net model.First,use the SE-Res Ne Xt-50 network for feature extraction,and use the extracted feature map as the input of Attention-Net to generate the attention map;Then,the feature map and the attention map are merged through the multiplication operation to obtain the mask,the attention map and the mask are global average pooling respectively,and divide the global average pooling result of the two;Finally,the result of the division is classified into five categories by the Softmax function.The model realizes the mutual promotion of channel attention and spatial attention through the combination of SE-Net and Attention-Net,so that the network’s attention is more focused on the pathological features of fundus images.And tested experimentally on the Eye PACE dataset,the model obtains a quadratic-weighted Kappa score of 0.7606,an ACA of 0.5574,and an average AUC value of 0.8719.(3)In order to realize the visual display of the classification results of diabetic retinopathy,a test platform of DR lesion detection system based on Web was built.The system has completed the closed-loop detection process: the front-end submits the fundus images,the server-side receives the images and calls the model for detection,and the Web page displays the classification results.The test results show that the system can intuitively display the degree of DR lesions in practical applications,which is helpful to improve the efficiency of doctors’ diagnosis.