Retinal vessel axis estimation through a multi-directional graph search approach

The analysis of blood vessels in images of retinal fundus is an important non-invasive procedure that allows early diagnosis and the effective monitoring of therapies in retinopathy. In order to derive a quantitative evaluation of the clinical features, such as vessel diameter and tortuosity, an accurate segmentation of the vessel network has to be performed. A new system for the automatic extraction of the vascular structure in retinal images is proposed. It is based on a sparse tracking technique via a multi-directional graph search approach. We consider the image as a weighted unoriented graph with arches connecting adjacent pixels and assume that vessels are minimum cost paths connecting remote nodes. An initial seed-finding algorithm based on fast 1- dimensional multi-scale matched filters is run over a regular grid. Simultaneous best-first search graph explorations start from each seed: when two search frontiers meet, the computed shortest path is recorded and exploited for a new search starting from it. New paths are found by iterating the procedure, until the entire vessel network is reconstructed. Lastly, in order to cover the unexplored region with lowcontrast vessels, a custom fixing procedure is run. 20 images have been used to test the algorithm, comparing the results with ground-truth manual segmentation. The method provides an average sensitivity of 96.2%.