Supervised, hysteresis-based segmentation of retinal images using the linear-classifier percentile

Image segmentation can be seen as a pattern classification problem, where each pixel is assigned, on the basis of, e.g., its gray level, either to the object or to the background class. In this setup, vessel segmentation is characterized by large class skew, as there are usually far more background pixels than vessel pixels and by weak separability, as there is a strong overlap between the two classes. The proposed hysteresis classification makes use of specific problem-domain knowledge to overcome such difficulties. We describe here a novel, supervised, hysteresis-based classification algorithm that we apply to the segmentation of retina photographies. This procedure is fast and achieves results that are superior to other vessel segmentation methods on similar data sets.