Comparative Study of Dermoscopic Hair Removal Methods

Abstract

When analyzing dermoscopic images, the hairs and their shadows on the skin may occlude relevant information about the lesion at the time of diagnosis. As far as we know, there is no method that quantitatively evaluates the performance of hair removal algorithms. In this work, we present a hair removal benchmark of six state-of-the-art algorithms, each with a different approach to segment and inpaint the hair pixels. To evaluate the algorithms, 13 dermoscopic images without hair were selected from the PH2 database. Next, two different hair simulators, providing hairs with a wide range of characteristics, are applied to these images. The results obtained with the hair removal algorithms on the simulated hair samples can be contrasted with the reference hairless images. To quantitatively assess their efficacy, we use a series of performance measures that evaluate the similarity between the original hairless image and the one obtained by each of the algorithms. Also, a statistical test is used to check the superiority of a method with respect to the others.