Research On Retinal Vessel And Optic Disc Segmentation

Human fundus tissues can be observed centrally in fundus pictures.Fundus pictures estimation and analysis is an important part in ophthalmic diagnosis.For one thing,fundus blood vessel analysis helps to diagnosis retinal disease,such as arteriosclerosis and diabetes.For another,fundus retinal optic disc analysis assistants to glaucoma and brain neuropathy diagnosis.It is helpful to ophthalmological diagnosis,treatment and screening by using image processing in fundus pictures assistant analysis.It is of widely range in practical value and application prospects.In recent years,many retinal vascular and optic disc segmentation methods have been emerging.However,most existing traditional machine learning vascular segmentation algorithms use single feature to train classifiers.And blood vessel segmentation with deep learning method depends on large fundus picture data sets.In addition,the performance of many optic disc segmentation methods is sensitive to the noise around the optic disc due to they are based on the whole retinal image.In view of the above,this paper proposed a retinal vessel segmentation method based on multi-feature fusion and a optic disc segmentation method based on faster regional convolution neural network and level set.The framework of this paper is as follows:To solve insufficient segmentation of small vessels caused by complex retinal vascular topology,this paper proposed a novel vascular segmentation algorithm.The algorithm integrates the linear structure,edge information,gray information and statistical information of retinal vessels to preserve the integrity of retinal vascular structure as much as possible.Firstly,linear feature,texture feature,moment feature,variance feature and gray feature of retinal vessels are extracted as vascular feature vectors.Secondly,random forest classifier is trained by fusion of various featuresFinally,a post-processing operator of vascular morphology is constructed to further improve the accuracy of fine vascular segmentation.The algorithm is tested on DRIVE and STARE datasets.Experiments show that the algorithm can ensure the connectivity of the results of vascular segmentation.The segmentation accuracy is good at the micro-vessels.The overall segmentation accuracy is consistent with most existing algorithms based on advanced machine learning.And it is universally higher than traditional retinal segmentation algorithms.To obviate the impact of background pixel in optic disc segmentation,this paper proposed an optic disc location and segmentation algorithm based on fast regional convolution neural network and level set.This method consists of two parts: rectangular optic disc area location and fine the optic disc boundary segmentation.In the localization part,the retinal features are extracted from the original retinal image convolution by ZF Net.Optic disc exact coordinate is obtained by means of regional convolution neural network.In the segmentation part,histogram matching and gray close operation are utilized to enhance the image of the optic disc region.Boundaries of optic disc are acquired though a level set method with shape constraint.The algorithm is tested on MESSIDOR dataset which are compared with different optic disc segmentation algorithms.The algorithm has the same disc coincidence rate as the state-of-the-art disc segmentation algorithm,and is robust to light and noise of fundus pictures.