Superpixel Inpainting For Self-Supervised Skin Lesion Segmentation From Dermoscopic Images

Automated and accurate segmentation of skin lesions based on dermoscopic images is an important task in clinical practice. However, limited labeled images and noisy annotations make the skin lesion segmentation task challenging. In this work, we propose a superpixel inpainting based self-supervised pretraining method to enhance skin lesion segmentation, the effectiveness of which is successfully identified on two public datasets. Superior performance on skin lesion segmentation is observed, with mean Jaccard indices of 76.5% and 84.3% being respectively obtained on the ISIC2017 and PH2 datasets.