Unsupervised Segmentation of Blood Vessels from Colour Retinal Fundus Images

Abstract

This paper represents an algorithm based on curvature evaluation and Entropy Filtering techniques with texture mapping for guideline. The method is used for the detection of blood vessels from colour retinal fundus images. In order to evaluate vessel-like patterns, segmentation is performed with respect to a precise model. We evaluate the curvature of blood vessels via carrying out eigenvalue analysis of Hessian matrix. This method allows to extract the fine retinal image ridge but introduces the effect of central light reflexes. We apply entropy filtering techniques to calculate the segmentations in relation to central reflex vessels. For efficient differentiation of vessels from analogous background patterns, we use spectral clustering to partition the image texture. It is an alternative of traditional intensity thresholding operation and allows more automatic processing of retinal vessel images. The detection algorithm that derives directly from this modeling is based on five steps: 1) image preprocessing; 2) curvature evaluation; 3) entropy filtering; 4) texture mapping; 5) morphology operation with application of vessel connectivity constraints.