Artifact Removal for Enhanced Melanoma Segmentation

We present a method for automatic lesion segmentation with removal of artifacts, defined as dark hair, dark corner, and color chart, before diagnosing melanoma. In order to detect the skin lesion more accurately, we apply our artifact removal method that uses morphological properties of artifacts for dermoscopic images. Subsequently, expectation-maximization is used to discriminate the skin lesion and normal skin. In our experiment, ISBI 2016 challenging dataset was used for validation of our method, and we achieved 90.6%, 91.2%, 93.4%, 78.7% in terms of accuracy, sensitivity, specificity, and Jaccard index for detection of the skin lesion.