The separation of retinal vessels is an important part of any automatic system for detection of vessel abnormalities caused by several systemic diseases such as diabetes, hyper tension and other cardiovascular conditions. It also provides a platform for medical researches on relation of various diseases and vessel characteristics, e.g. Arteriolar-to-Venular Ratio (AVR) and vessel tortuosity. This paper presents a new method for automatic separation of retinal vessels based on the analysis of a graph extracted from retinal vasculature.

In the proposed method, we first use vessel segmented image to extract the vessel centerlines. The centerline network is then modeled as a directed graph where each node represents an intersection point in the vascular tree, and each link corresponds to a vessel segment between two intersection points. Link directions are based on the vessel growth direction in the retina. In the next step, nodes and links inside the optic disk are eliminated and graph traversal is started from the nodes on the optic disk border. Each input link is given a unique label and the labels are propagated through the graph according to the node type which can be a bifurcation or a crossover. In bifurcations, the input and all output links must have the same labels, whereas in crossovers, input links are assigned different labels and each output link is connected to the most similar input in the orientation, width and color.

This paper proposed a new method for automatic separation of distinct retinal vessels. The separation algorithm was tested on the DRIVE dataset. The result is compared with the manual labeling of the images and the accuracy value of 0.65 is obtained. Fig.1 shows the result with distinct vessels depicted in different colors.

The automatic algorithm for separation of distinct retinal vessels will help the experts to analyze retinal images and detect the abnormalities of vascular topology in less time. It also can be used as the initial step of an automatic vessel classification system which is worthwhile for automatic screening programs.