Blood Vessel Segmentation In Pathological Retinal Image Based On SVM

With the development of medical imaging, more and more medical images are applied to clinical diagnose and research. Retinal image is increasingly important in recent years. Retinal image processing contains retinal image recognition, retinal vessel segmentation, optic disk detection, fovea detection and pathological areas detection. Blood vessels are the main and stable structure in retinal image. Moreover, retinal vessel segmentation is the fundamental work for retinal image processing. Information of retinal blood vessel is helpful for diagnose of many diseases. Based on the stable position of retinal blood vessel, detections of other structures in retinal image are easier and more effective. Retinal blood vessel is unique for everyone and it is hard to fake, so that recognition methods based on it is safer than many other methods.In this paper, the method is focus on the challenges on segmentation of retinal blood vessel described below. Firstly, uneven illumination images make segmentation more difficult. Secondly, Edges of other structures, such as optic disk, fovea and pigmented epitheliums, are often error detected as vessels. Besides, in pathological retinal image, the biggest challenge is distinguish of the edges of pathological areas and blood vessel. Few segmentation methods, especially supervised methods, are designed for pathological retinal image. Feature extraction algorithm in literature is not enough for the blood vessel segmentation problem in pathological retinal image.All the methods in literature about retinal blood segmentation roughly contain two categories:supervised methods and unsupervised methods. Supervised methods obtain segmentation results with labeled information whereas unsupervised methods do not need labeled information. Unsupervised methods mainly contain four categories: matched filtering, vessel tracking, morphology processing and model based algorithms. Supervise method mainly contains two parts:feature extraction and classification. Many unsupervised methods are designed for pathological retinal image and have many limitations. However, few supervised methods are designed for pathological retinal image. In this work, a supervised method is proposed and has a good performance on this problem.New effective feature extraction algorithms are designed for the segmentation of blood vessel in pathological retinal image. Extract features in a local area solve the uneven illuminate problem. New features contain shape-based line operators, the distribution of gray levels and multi-scale morphology operations and contain more information of blood vessels and background. Support vector classifier is applied in this work. The performance of the algorithm is tested on STARE (Structured Analysis of the Retina) and DRIVE (Digital Retinal Images for Vessel Extraction) database. The proposed method has good performance both on health and pathological retinal image and is robust for uneven illuminate images. Besides, the method has good application ability.