Research On Diabetic Retinopathy Grading And Fundus Blood Vessel Segmentation Based On Convolutional Neural Network

With the rapid development of deep learning,the use of deep learning methods to assist doctors in disease diagnosis and treatment has become a hot research direction.The eye is a high incidence area of diseases,so the use of deep learning methods to diagnose eye diseases is an area that urgently needs in-depth research.Due to changes in living environment and conditions,more and more people are suffering from diabetes mellitus.With the increase in the duration of diabetes mellitus,patients may also suffer from the complication of diabetic retinopathy,which is extremely harmful to eye health.If diagnosis and treatment are not performed in time,the best time for treatment will be delayed,resulting in impaired vision or even blindness.Secondly,fundus blood vessels are an important indicator and basis for the diagnosis and treatment of many eye diseases.So it is very necessary to use deep learning methods to segment the fundus blood vessels to assist doctors in the diagnosis and treatment of related eye diseases,which greatly reduces The burden of doctors and the efficiency of diagnosis are improved.Therefore,it is very practical to design related deep learning solutions for these two works,which can help prevent eye diseases and improve the efficiency of doctors’ diagnosis and treatment.To address the problems of diabetic retinopathy grading and fundus blood vessel segmentation,this paper proposes a diabetic retinopathy grading scheme based on the Noisy-student self-training method and the DP-Efficient Net classification model and the fundus blood vessels segmentation model DA-Res2 UNet,respectively,based on convolutional neural networks.The main work in this paper is as follows:（1）For the diabetic retinopathy grading task,considering that the diabetic retinal data set is generally small,the model is very prone to overfitting to affect its performance.However,there are a lot of unlabeled medical data in the real environment,so this scheme uses the Noisy-student self-training method based on semi-supervised learning to use these unlabeled data to improve the performance of the classification model,and finally achieve diabetic retinopathy classification through the DP-Efficient Net.Subsequent experiments have fully proved the effectiveness of this diabetic retinopathy grading scheme.（2）According to the characteristics of the fundus blood vessel segmentation task,this paper proposes the DA-Res2 UNet model.This model is based on the classic medical image segmentation model UNet,and replaces the ordinary convolution module with the Res2 module to help the model obtain multi-scale information;Secondly,this model adds the Dual attention module,which helps the model to obtain global dependency information to improve the segmentation accuracy;In order to prevent the model from overfitting,the segmentation model adds the Dropblock module,which is more suitable for the convolutional layer.Subsequent experiments have also demonstrated the validity and feasibility of the DA-Res2 UNet.