Research On Retinal Blood Vessel Segmentation And Arteriovenous Classification Based On Improved U-Net

Studies have shown that diabetic complications can cause abnormalities in retinal vessel morphology.Clinically,the morphological changes of retinal artery and vein in a certain range in fundus image are used as the basis for judging the course of diabetic retinopathy.However,doctors manually perform multi-point sampling and measuring retinal vessel in clinical practice.This method is time-consuming,laborious and subjective,which is not suitable for a large number of screening work of diabetic retinopathy.Therefore,the realization of automatic segmentation and automatic classification of artery and vein of retinal vessel in fundus image by deep learning is of great significance for the objective and rapid diagnosis of early diabetic retinopathy.In view of U-Net's excellent performance in the field of medical image analysis,therefore,U-Net framework is used to achieve the segmentation and the classification of artery/vein of retinal vessel in fundus image respectively.In the part of retinal vessel segmentation,three kinds of U-Net derived networks through experiments were compared and analyzed.Due to the iterative architecture of Iter Net adds fine module mini-UNets after the basic module U-Net,it has a deeper understanding of retinal vessel network and can connect retinal vessel segments together,which solves the problem of poor vessel connectivity in segmentation results.Therefore,Iter Net is used as the retinal vessel segmentation model in this thesis.In the retinal vessel artery/vein classification part,aiming at the unbalance between the numbers of pixels on retinal artery or retinal vein and the number of background pixels in the fundus image,and the poor segmentation ability of existing methods,this thesis proposes a retinal vessel artery/vein classification method based on the multiply Iter Net network.The fundus image and the segmented retinal vessel image are used as input,which reduces the deviation of the label distribution.And this method divides the retinal vessel artery/vein classification task into two parts: retinal artery segmentation and retinal vein segmentation.The relatively easy vessel segmentation task is used to improve the performance of retinal vessel artery/vein classification.Finally,experiments are carried out on the public datasets and a private dataset.The experimental results show that on the DRIVE dataset,the accuracy of retinal vessel segmentation of the method in this thesis is 0.9556,the accuracy of artery/vein classification on the center line of retinal vessel is 0.9091,and the accuracy of artery/vein classification on the center line of main retinal vessel(width greater than 2pixels)is 0.9347;on the STARE dataset,the accuracy of retinal vessel segmentation is0.9641;on the private dataset,there are good results of retinal vessel segmentation and artery/vein classification.The method in this thesis can achieve good retinal vessel automatic segmentation and artery/vein automatic classification in fundus image.