Classify in vivo confocal microscopy corneal images by tortuosity is complicated by the presence of variable numbers of fibres of different tortuosity level. Instead of designing a function combining manually selected features into a single coefficient, as done in the literature, we proposed a supervised approach which selects automatically the most relevant combination of shape features from a pre-dened dictionary. To our best knowledge, we are the first to consider features at different spatial scales and show experimentally their relevance in tortuosity modelling. Experimental results using a data set of 90 images provided by our clinical collaborators at the Harvard Medical School and Massachusetts Eye and Ear Infirmary show that our framework yields an accuracy indistinguishable, overall, from that of experts when compared against each other.
Roberto Annunziata – Progress Report 2
//