Application Research Of Diabetic Retinopathy Based On Deep Learning

Diabetic retinopathy is a common eye complication that endangers the vision health of diabetic patients.Conventional methods for the detection of diabetic reticulum lesions have shortcomings such as unstable accuracy or allergic reactions.Based on deep learning technology,this thsis proposes a fundus retinal blood vessel segmentation model,a classification model for the recognition and classification of glucoreticular lesions,moreover,realizes the application of automatic detection of glucoreticular lesions.The purpose of this study is to use computer technology to assist doctors in making efficient,what’s more,accurate diagnosis of patients.The main contents are as follows.This thsis proposes a retinal blood vessel segmentation model based on DCHAU-Net,which mainly improves the feature extraction,pooling together with skip connection modules in the conventional U-Net structure.The specific improvement points are:(1)For the traditional volume Because of the problem that it is difficult to extract the pixel features of blood vessels after unknown deformation,a deformable convolution module is proposed,which can effectively adapt to the changing target features;(2)For the pooling operation in the U-Net encoder It is easy to miss the problem of image detail information.It is proposed to use a dilated convolution module in the compression path.This structure can increase the receptive field area of the feature map while keeping the image size unchanged;(3)For the encoder with decoder in order to solve the problem of noise interference when perfovrming feature splicing,an attention threshold network is proposed to add an attention gate network to the skip link,which can enhance the blood vessel pixel features by suppressing the background region.The proposed DCHAU-Net model is tested on the public fundus dataset to evaluate the model performance,the network accuracy is 0.9555,the sensitivity is0.8291,the specificity is 0.9740,the AUC value is 0.9805 on the DRIVE dataset.Comparing the model in this thsis with other models,it is concluded that the improved DCHAU-Net model proposed in this thsis has a better segmentation effect on the retinal blood vessels of the fundus.This thsis proposes an improved Res Ne St based diabetic retinopathy lesion recognition classification model.The model is mainly improved on the attention module and loss function of the original Res Ne St model.The specific improvement points are:(1)The original Res Ne St network has insufficient feature expression ability For this problem,an attention mechanism that combines the soft attention algorithm of spatial-channel mixing with the deep connection attention algorithm is proposed.This structure can connect adjacent attentions to improve the performance of the soft attention model;(2)Aiming at the problem of unbalanced number of sample categories in the dataset,a focal loss function is proposed to increase the weight of difficult-to-classify samples to improve the classification effect of the model for different lesion categories.The performance of the improved Res Ne St model in this thsis is evaluated on the public EYEPACS 2015 dataset,the results show that the model has an accuracy of 85.71%,a sensitivity of 86.23%,a specificity of 72.97% in the DR five-classification task.Comparing the model in this thsis with other algorithms,it is concluded that the improved Res Ne St model proposed in this thsis has a better classification effect on the diabetic reticulum lesions.This thsis designs and implements a front-end with back-end interactive automatic detection application of diabetic retinopathy lesions.The system is based on the mainstream lightweight back-end Flask framework with mature front-end page technology,also realizes the Web-based diabetic retinopathy disease intelligent detection application v1.0 version,the diabetic fundus image detection system greatly improves the detection efficiency of doctors.The research results of this thsis have exceedingly important application value for improving the efficiency of doctors’ diagnosis,reducing the blindness rate of patients,rationally distributing medical resources as well as protecting the vision health of diabetic patients.