Dermoscopic interest point detector and descriptor

Abstract

Dermoscopy is an imaging technique dermatologists use to better visualize pigmented skin lesions (PSLs) and determine their malignancy. Dermoscopic features revealed by this technique have been shown to correlate with histopathology features, and are used as diagnosis indicators by many dermatologists. Hence, automated detection and classification of these features is the first step toward computer-aided diagnosis of melanoma in dermoscopy. In this paper, we present a novel scale- and rotation-invariant feature detector and descriptor specifically designed as a general visual vocabulary of dermoscopic features. We compare our feature detector and descriptor to the popular interest point detectors in the vision community, namely, SIFT, and a more recent fast variant, SURF. We demonstrate that our feature detector is more discriminative and reliable for dermoscopic features.