Unsupervised Recognition of Retinal Vascular Junction Points

Landmark points in retinal images can be used to create a graph representation to understand and to diagnose not only different pathologies of the eye, but also a variety of more general diseases. Aim of this paper is the description of a non-supervised methodology to distinguish between bifurcations and crossings of the retinal vessels, which can be used in differentiating between arteries and veins. A thinned representation of the binarized image, is used to identify pixels with three or more neighbors. Junction points are classified into bifurcations or crossovers according to their geometrical and topological properties. The proposed approach is successfully compared with the state-of-the-art methods with the benchmarks DRIVE and STARE. The recall, precision and F-score average detection values are 91.5%, 88.8% and 89.8% respectively.