Abstract

The segmentation of the vascular network from retinal fundal images is a fundamental step in the analysis of the retina, and may be used for a number of purposes, including diagnosis of diabetic retinopathy. However, due to the variability of retinal images segmentation is difficult, particularly with images of diseased retina which include significant distractors.
This paper introduces a non-linear filter for vascular segmentation, which is particularly robust against such distractors. We demonstrate results on the publicly-available STARE dataset, superior to Stare’s performance, with 57.2% of the vascular network (by length) successfully located, with 97.2% positive predictive value measured by vessel length, compared with 57% and 92.2% for Stare. The filter is also simple and computationally efficient.