Vessel Segmentation In Retinal Image Based On Radial Projection And Semi-supervised Learning Approach

Retinal image processing and analysis are hot issues in medical and biometrics re-search, which have important value in theoretical research and extensive application back-ground. In medical research, retinal images provide considerable information on patholog-ical changes which may reveal some disease. It is believed that a large number of patients can be prevented, in part through periodic screening and early diagnosis. Identification of retinopathy is mainly done by ophthalmologists' manual analysis; however, it takes ophthal-mologists an expensive time. Thus, an automatically computerized approach can be used to improve the efficiency of assessment of fundus images.Reliable vessel extraction is a prerequisite for subsequent retinal image analysis and processing because vessels are the predominant and most stable structures appearing in the images. There are number of good ways to detect blood vessels in retinal image proposed by researchers. But most of them trend to lose thin vessels in the low contrast background. We investigate an automated method in detection of the centerline to solve the problem. Each pixel in the filed of view is performed to obtain a projected curve by radial projection. By making an analysis of the peak property of the curve, we can judge whether the pixel is a vessel candidate or not. Then a removal scheme is used to eliminate false candidate and get the true centerline candidate. The centerlines include thin vessels with low contrast.Steerable filters and semi-supervised approach are introduced to exact the major struc-ture of retinal blood vessels. We employ steerable filters to adaptively enhance blood vessels in retinal image, then vector features are constructed to describe the pixel using line operator under different scale. The initial SVM classifier is generated by training the manual labeled segmentation. We use self-training to mark unlabeled data as labeled data and add them into the training, to improve the performance of classifier. The input retinal image is classified as two classes (vessel and non-vessel pixels), and major structure of vessel segmentation is obtained.After the union of the centerlines and the major structures of retinal blood vessels and noise removal scheme, we use morphological operation to obtain the final vessel segmentation. The performance of the proposed method is evaluated and compared with other approaches on two publicly available DRIVE and STARE databases. The proposed method is able to extract thin vessels with low contrast and better big vessels and improve the accuracy of segmentation of vessels.