Domain prior based superpixel propagation for optic cup localization

In this paper, we present an unsupervised framework using domain priors extracted from the primary structures of the optic nerve head for automated optic cup localization. Our approach provides 3 major contributions. First, we identify a new domain prior, optic cup origin. This prior is derived from the physiological understanding that the central retinal vessels traces its origin from the optic cup before extending to the rest of the retinal. Second, we propose extracting the features of the optic nerve head from superpixels, which are obtained from low-level grouping and have more natural and descriptive features than pixel based techniques. Third, the domain knowledge comprising of optic cup origin and cup pallor, and the extracted features from superpixels are then used to drive a similarity-based label propagation and refinement scheme for the optic cup localization. Our approach was validated on a clinical online dataset, ORIGA-light, of 650 population-based images. Overall, our approach is able to achieve a 32.2% nonoverlap ratio (m1), a 33.8% relative absolute area difference (m2) and a 10.6% absolute CDR error (δ).