Melanoma classification from Hidden Markov Tree features

Melanoma detection relies on visual inspection of skin samples under the microscope via a qualitative set of indicators, causing large discordance among pathologists. New developments in pump-probe imaging enable the extraction of melanin intensity levels from skin samples and provide baseline qualitative figures for melanoma detection and classification. However, such basic figures do not capture the diverse types of cellular structure that distinguish different stages of melanoma. In this paper, we propose an initial approach for feature extraction for classification purposes via Hidden Markov Tree models trained on skin sample melanin intensity images. Our experimental results show that the proposed features provide a mathematical microscope that is able to better discriminate cellular structure, enabling successful classification of skin samples that are mislabeled when the baseline melanin intensity qualitative figures are used.