Self-distillation with Edge-Aware Pseudo-Masks for Skin Lesion Segmentation

Label scarcity remains a significant challenge in skin lesion segmentation due to the cost and expertise required for manual annotation. Self-supervised learning (SSL), particularly self-distillation, offers a promising solution by enabling representation learning without labeled data through a student–teacher framework. We propose a DINO-style self-distillation framework enhanced with edge-aware pseudo-masks generated by the Random Walker algorithm after modified DullRazor preprocessing. Unlike prior works that rely solely on either distillation or pseudo-labeling, our approach jointly optimizes a distillation loss and an auxiliary segmentation loss, encouraging the encoder to capture both semantic structure and boundary-sensitive features. Experiments on the ISIC 2018 dataset show that our method outperforms supervised baselines and other SSL approaches under extreme label scarcity (0.2–1%), achieving up to +15.5% Jaccard and +13.2% F1-score improvement at 0.2% labels. These results demonstrate that edge-aware pseudo-mask guidance can significantly boost representation quality and enable clinically useful segmentation performance with minimal annotation effort.