Statistical pattern analysis of blood vessel features on retina images and its application to blood vessel mapping algorithms

Computer based modeling and analysis of blood vessel (BV) networks is essential for automated detection and tracking of anomalies and structural changes in retina images. Among many published techniques for automated BV mapping, optimal selection of thresholds to delineate BV pixels from their background pixels remains an open problem. In this paper we propose a novel representation of a BV pixel feature, daisy graph, using rotational contrast transform (RCT), and two feature descriptors energy E_p and symmetry difference S_p of the daisy graph. Non-BV pixels are separated from BV and boundary pixels based on E_p. Fitness of the lognormal distribution to S_p of BV pixels with negative E_p has been tested extensively for images in the STARE and DRIVE databases. Based on statistical pattern analysis in the feature space, we propose a fast self-calibrated BV mapping algorithm which achieve comparable and statistically sound performance as contemporary solutions.