Segmentation of optic nerve head images

Abstract

Segmentation of the optic nerve head or optic disc in digital retinal fundus photographs is a non-invasive procedure that plays an important role in early detection of abnormalities of the eyes, particularly glaucoma diseases. Developing an automatic system, we employ image processing techniques coupled with graph cut algorithms from combinatorial optimization. Avoiding the need of manual pre-segmentation for constructing an initial graph, a supervised learning approach is effectively and efficiently applied. Crucial information is extracted from a set of labeled binary masks and integrated into weight assignments of the edges of the graph. We associate the characteristically bell-shape of a Gaussian distribution with the rounded circular-shape of the optic disc. Our approach was validated and evaluated on the RIM-ONE open database. Segmentation is successful on 91.12% of the entire 169 images, achieving 91% sensitivity and 88% accuracy.