The Research On Aided Diagnosis Of Diabetic Retinopathy Based On Attention Model

Diabetic retinopathy is a kind of fundus disease caused by diabetes,which mainly occurs in retinal vessels.At present,the diagnosis of diabetic retinopathy relies on the analysis of the pathological characteristics of retinal image by ophthalmologists.However,the number of patients is large and the time of artificial screening is relatively long,which makes the patients’ condition worsen due to missing the best opportunity of diagnosis and treatment,and result a great threat to blindness.With the rapid development of deep learning in the medical field,the use of deep learning method combined with computer-aided diagnosis technology can improve the timeliness of diabetic retinopathy detection to a certain extent and shorten the diagnosis time of patients,which has a broad clinical application prospect for the construction of diabetic retinopathy intelligent diagnosis system.The main work of this paper is to study the assistant diagnosis of diabetic retinopathy based on various attention models for different detection tasks.Based on the ordinary deep learning network fusion attention model,to achieve the classification of diabetic retinopathy,retinal vascular segmentation and detection of focal points of diabetic retinopathy and other tasks.Through experiments on different standard datasets,the feasibility and superiority of this algorithm in fundus image analysis are proved,which provides an effective reference for the assistant diagnosis and treatment of diabetic retinopathy.The main research work of this paper includes the following aspects:1.In the task of diabetic retinopathy classification,due to the poor ability of the common convolutional neural network to extract the focus features,which is not enough to support the needs of the task of diabetic retinopathy classification,this paper designs a deep neural network based on channel domain attention model(Visual Geometry Group Network with Squeeze and Excitation Networks,SEVGG)for the classification of diabetic retinopathy.Experiments are carried out on two kinds of standard data,from the perspective of medical image evaluation indicators,to evaluate the results of grading of diabetic retinopathy under different network structures,verify the effectiveness of the algorithm;at the same time,a series of visual experiments are carried out to enhance the effect of channel domain attention model,and the experimental results prove the feasibility and superiority of the algorithm in this paper.2.In order to effectively solve the problem of poor retinal blood vessel segmentation,an attention enhancement model based on convolutional block(Convolutional Block Attention Module for U-net,CBAM-Unet)is improved for segmentation of retinal blood vessels.First,preprocess the fundus image to highlight the characteristics of retinal blood vessels;then train the network model,and use the trained model to segment and predict the retinal blood vessels.By testing on two standard fundus data sets,both have achieved good experimental results and achieved effective segmentation of retinal blood vessels.3.Because the segmentation effect of the existing algorithm is not obvious and the detection steps are tedious,it is difficult to support the practical application needs,so this paper constructs a multi-scale segmentation network for the detection of diabetic retinopathy focus points(Deep Multiscale Segmentation Network with Selective Kernel Networks for Diabetic Retinopathy,DMSSKN)to perform the task of diabetic retinopathy focus points segmentation.By using the attention model with selective convolution kernel,the feature information of the focus on different scales is extracted,which can effectively segment the focus and improve the segmentation accuracy.