Melanoma is one of the most dangerous types of skin cancer and causes thousands of deaths worldwide each year. Recently dermoscopic imaging systems have been widely used as a diagnostic tool for melanoma detection. The first step in the automatic analysis of dermoscopy images is the lesion segmentation. In this article, a novel method for skin lesion segmentation that could be applied to a variety of images with different properties and deficiencies is proposed. After a multi-step preprocessing phase (hair removal and illumination correction), a supervised saliency map construction method is used to obtain an initial guess of lesion location. The construction of the saliency map is based on a random forest regressor that takes a vector of regional image features and return a saliency score based on them. This regressor is trained in a multi-level manner based on 2000 training data provided in ISIC2017 melanoma recognition challenge. In addition to obtaining an initial contour of lesion, the output saliency map can be used as a speed function alongside with image gradient to derive the initial contour toward the lesion boundary using a propagation model. The proposed algorithm has been tested on the ISIC2017 training, validation and test datasets, and gained high values for evaluation metrics.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)