Research On Segmentation Algorithm Of Retinal Vessels

The analysis of the characteristics of retinal blood vessels can help to detect and treat diseases in the early stage,such as cardiovascular disease and diabetes.Since cardiovascular diseases and ophthalmopathy have a serious impact on human life,the analysis of retinal blood vessels is of great clinical significance in revealing important information of systemic diseases and supporting diagnosis and treatment.There are many segmentation methods of retinal fundus vessels,including traditional feature extraction and deep learning based segmentation methods.The traditional segmentation method of feature extraction takes advantage of the characteristics of the vascular pixels in retinal fundus images,constructs the feature extraction operator,and uses the threshold value to complete the segmentation.Because there are a large number of small vascular pixels in fundus images,and the identification of these vessels and their surroundings is low,the simple,especially the single feature extraction operator cannot detect these small vascular pixels,resulting in low precision of segmentation.Although the development of deep learning theory in a certain extent,improved the precision of retinal blood vessels image segmentation,but fundus retinal blood vessels have very strong tree unstructured characteristics,how can extract more fully unstructured characteristics of blood vessel,improve the connectivity of vessels,thus improve the blood vessels segmentation accuracy become the key problems of retinal blood vessels segmentation.In addition,the optic disc region generated in the process of retinal fundus image acquisition is an important factor affecting the vascular segmentation of retinal images.How to segment the image of the fundus vessel by the light spot in the descending optic disc region has always been a hot issue in the segmentation of the fundus vessel.In order to solve these problems,two segmentation methods,traditional feature extraction and deep learning,were used to conduct in-depth research on fundus retinal images in this study.Aiming at the problem of low resolution of terminal branch of blood vessels in fundus retinal image,this paper proposes a new traditional segmentation integration schemes,it combines local pixel features,morphological transform and matched filtering,composed of3 d data,graph cut method is used for 3d features are classified,thus realize the limited accuracy of segmentation for retinal fundus images.In the existing deep learning-based segmentation framework,the convolution kernel is mostly based on the image data with obvious structured data for feature extraction.In order to more fully extract the non-structural features of blood vessels and increase the connectivity of segmentation blood vessels,this paper proposed a new segmentation framework,the framework using graph convolution kernels instead of the traditional segmentation convolution operation,the structured obvious image data abstraction for data,based on this,The image convolutional neural network is used to extract the features of the data and further achieve the purpose of retinal fundus image segmentation.Aiming at the problem that the identification of some blood vessel pixels and surrounding areas is low,and there is a bright spot in the optic disc area of retinal fundus image.Based on this,a new neural network model,called U-Capsule model,is proposed in this paper.It acts locally on capsule network and fuses multi-scale information of U-NET network.The model separated the optic disc and non-optic disc according to the gray level,and then accomplished the high-precision segmentation of retinal fundus images.At the same time,based on the segmentation results and the idea of block matching,the noise reduction of the highlighted region of optic disc was completed.