The Research On Vessel Segmentation And Classification Of Diabetes Retinopathy Based On Retinal Fundus Image

With the continuous development,deep learning has significant effect in image segmentation and classification.Using deep learning technology,computer-aided diagnosis of ocular diseases has been widely studied and concerned through retinal fundus images.Retinal blood vessels are a critical step in the accurate diagnosis,early treatment,and surgical planning of ocular diseases.Many blinding eye diseases can be directly observed from the retinal blood vessels of the fundus images.Early treatment of diabetic retinopathy can effectively delay or avoid the progression of visual impairment.Therefore,it is of great clinical significance to realize the classification of blood vessels under the retinal fundus image and the classification of diabetic eye diseases.In view of the current problems of vessel segmentation and diabetic retinopathy on retinal fundus image,this paper proposes a retinal fundus vessel segmentation method based on Weighted Res-UNet and a classification and visualization method of diabetic retinopathy based on Grad-CAM.The main research work and innovations of this paper are as follows:(1)In view of the difficulty of blood vessel segmentation on the retinal fundus image,the Weighted Res-UNet retinal vessel segmentation method is proposed.The method uses adaptive histogram equalization(CLAHE)and conventional image enhancement for data preprocessing.Our method uses the U-NET network,Adding the weighted Attention mechanism and the skip connection method to make important improvements.It can learn more about the characteristics of vessel and non-vessel pixels,and better maintain the retinal vessel tree structure.Experiments on the DRIVE and STARE datasets yielded better segmentation results than most existing segmentation methods,achieving accurate and robust segmentation results.(2)In view of the problem of poor generalization of traditional DR classification methods and weak interpretability of DR classification network based on Deep learning methods,we proposed a grad-CAM-based classification and visualization method for diabetic retinopathy.Our method improve the Inception-V3 model and use Grad-CAM network visualization algorithm to interpret the classification results and find potential lesion areas.Then we cluster the features of normal and potential lesion areas by codebook and calculate the BoW feature histogram vectors.Finally,the final DR classfication is carried out by multi-layer neural network.The experimental results on EyePACS dataset show that the proposed classification and visualization method achieves good classification results.In summary,in view of the difficulties of small and thin vessels segmentation in retinal fundus image,the poor segmentation effect in low contrast areas such as optic disc area,and the difficulty of maintaining retinal vessel tree structure in existing segmentation results,this paper proposes Weighted Res-UNet retinal vessel segmentation method.Aiming at the poor generalization of traditional DR classification methods,and DR classification network has weak explanability based on deep learning.A Grad-CAM-based classification and visualization method for diabetic retinopathy is proposed.The experimental results verify the validity of the proposed method,which provides a strong guarantee for the follow-up accurate diagnosis,early treatment and surgical planning of eye diseases.