Classification of Arteries and Veins in Retinal Images using Vessel Profile Features

In this work an automated method is introduced to distinguish arteries from veins in eye fundus images. The main challenges of this task are the similarity of the two vessel types and the vast variability in different images. We assume that the vasculature is already extracted and represented by vessel segments. Based on local image features, vessel profile characteristics are extracted and used for clustering the major vessels near the optic disc. The presented method is tested on the MARS database including 448 retinal images with ground truth (artery and vein labels). The achieved results are promising: for 80% of the images the classification error is less than 30%. These results can be further improved by structural analysis of the vasculature.