Segmentation of retinal blood vessel based on Radon Fourier Transform and Gaussian Process

Abstract

Diabetic retinopathy is one of the most common cause of blindness and it can be prevented if detected in early stages. Accurate segmentation of retinal vessels is momentous for pathological analysis and plays a decisive role in early diagnosis of diabetes. Accordingly, the extraction of the centerlines of tubular objects is an invaluable task to be achieved. In previous studies, assorted methods have been presented for retinal blood vessel segmentation. In this research, we propose a new approach to track the vessel and implement the segmentation based on Gaussian Process and Radon Fourier Transform. The curvature through vessel is assumed as a Gaussian Process with zero mean. Specific features are extracted by Radon Fourier Transform to be considered as inputs of Gaussian Process. As a final remark, the kernelized covariance matrix is learned from training data to estimate the vessel direction. Moreover, the bifurcations are Determined using multiple GPs. Since the Radon Transform is robust against noise, proposed algorithm is noise robust too, and performs well in comparison to the other tested methods. The results demonstrate that the proposed method successfully deals with the bifurcations, and tracks the thin vessels with a high accuracy. The algorithm can perform quite well on a variety of tubular objects. Consequently, it can be developed for other researches as future work. Comparisons are applied on the publicly available DRIVE and STARE databases.