Automatic Extraction of Blood Vessels, Bifurcations and End Points in the Retinal Vascular Tree

In this paper we present an effective algorithm for automated extraction of the vascular tree in retinal images, including bifurcations, crossovers and end-points detection. Correct identification of these features in the ocular fundus helps the diagnosis of important systematic diseases, such as diabetes and hypertension. The pre-processing consists in artefacts removal based on anisotropic diffusion filter. Then a matched filter is applied to enhance blood vessels. The filter uses a full adaptive kernel because each vessel has a proper orientation and thickness. The kernel of the filter needs to be rotated for all possible directions. As a consequence, a suitable kernel has been designed to match this requirement. The maximum filter response is retained for each pixel and the contrast is increased again to make easier the next step. A threshold operator is applied to obtain a binary image of the vascular tree. Finally, a length filter produces a clean and complete vascular tree structure by removing isolated pixels, using the concept of connected pixels labelling. Once the binary image of vascular tree is obtained, we detect vascular bifurcations, crossovers and end points using a cross correlation based method. We measured the algorithm performance evaluating the area under the ROC curve computed comparing the number of blood vessels recognized using our approach with those labelled manually in the dataset provided by the Drive database. This curve is used also for threshold tuning.