Automatic refinement of vascular tracking in retinal images: False vessels detection

Abstract

A reliable vessel extraction is often a prerequisite for any retinal image analysis. A postprocessing step for the automatic detection of false vessels is then necessary, mainly in bad quality or "difficult" images (e.g. because of lesions or hemorrhages). The proposed method considers the vessels as described by their centerline pixels and for each transversal section it classifies pixels as vessel or no-vessel using the fuzzy C-means technique. A vessel is classified as true if the distances between the edges and the centerline follow a monomodal gaussian distribution with a small standard deviation, i.e. if the they are parallel and symmetrical in respect to the centerline. Otherwise, if the distribution is monomodal but with a large standard deviation or bimodal, the vessel is recognized as false. The algorithm showed good performance both in the training and testing dataset, containing both good and bad quality images, with Auc of ROC curves always larger than 0.9.