Retinal Vessel Segmentation Based On Encoder-Decoder Structure

Retinal blood vessel segmentation is an important step in the screening of fundus diseases.It has important value both in clinical research and early diagnosis of diseases.However,these vessels in retina image show complex structures,causing difficulties in manual segmentation.How to effectively segment the retinal vessels is still a hot and difficult point of current research.Aiming at the mentioned task,this thesis proposes two retinal vessel segmentation algorithms based on encoder-decoder structure by using deep learning.The main research work of this thesis is as follows:(1)A multi-scale segmentation method based on wide activation is proposed.This method focuses on the "width" of deep learning.Firstly,it takes wide activation as the research basis and adds the optimization design of joint regularization.Wide activation changes the width of the neural network,and joint regularization optimizes the network in two aspects: the data level and the training level,which speed up the network convergence while improving the segmentation performance.Secondly,in order to obtain more tiny blood vessels,a residual atrous spatial pyramid module is designed to capture multi-scale information.Experiments show that the proposed method,which uses fewer parameters,captures small blood vessels effectively while ensuring speed and accuracy.(2)A segmentation method combined bi-directional convolutional LSTM with iterative encoder-decoder network is proposed.Aiming at the problem of insufficient continuity in the segmented vessels of the retinal image,this method,which takes an iterative encoder-decoder network as the basic framework,uses a dense feedforward cascade method for cross-layer connections,therefore,the ability to learn the continuity of vessel branches is guaranteed.In order to strengthen the model’s ability to predict blood vessels,this thesis uses bi-directional convolution LSTM during the cross-layer connection process,thus the temporal and spatial characteristics of the retinal image are obtained and the segmentation result is optimized.Before training,data augmentation is used in the network to expand the samples to further improve the generalization performance.Experimental results show that the proposed method improves the segmentation performance,and the segmented blood vessels are more continuous.(3)This thesis conducted experiments on two fundus image data sets,DRIVE and STARE.In order to evaluate the mentioned two algorithms comprehensively,three comparison experiments are designed,including comparison of different algorithms within the data set,cross-validation between data sets,and the impact of improvements.All the results show that both methods are robust,and the segmentation performance is competitive.