Learning-based segmentation of optic disc in retinal images using clustering trees and local mode filtering

Abstract

Delineation of the optic disc boundary in retinal images is the first step towards the computation of cup-to-disc ratio, an important indicator of ophthalmic pathologies such as glaucoma. This paper proposes the combination of learning-based clustering trees with local mode filtering for the segmentation of the optic disc region in retinal images. The algorithm identifies candidate optic disc region by extracting and pooling low-level features at different clustering resolutions from the filtered region-of-interest in two color channels. Thereafter, we use learned geometric properties such as area, eccentricity and solidity to extract high-level features for the identification of connected components, which most likely belong to the optic disc region. The final stage pools and fully connects these connected components into a single segmented region. Performance evaluation on three publicly available datasets from IDRID, DRISHTI-GS and MESSIDOR demonstrate promising results that are comparable to state-of-the-art algorithms.