Authors - Maulikkumar Pandya Abstract - Skin lesion segmentation is essential for computer-aided dermatological diagnosis, but reliable pixel-level annotations are costly and require experts. To reduce dependence on manual labeling, pseudolabeling combined with foundation models such as the Segment Anything Model (SAM) has been explored; however, most pipelines rely on a single pseudo-label per image, which can introduce boundary bias when pseudo-labels are noisy. In this paper, we compare two U-Net training pipelines built on pseudo-labels generated using U²-Net and SAM. The first pipeline follows a single pseudo-label inheritance strategy as a strong annotation-free baseline. The second pipeline synthesizes multi-style pseudo-labels (tight/moderate/loose) and applies agreement-based learning to supervise only high-confidence consensus regions while suppressing uncertain boundary pixels. No ground-truth masks are used during training; manual annotations, when available, are used only for offline evaluation. Experiments on ISIC 2018 under a pseudo-reference protocol show improved boundary behavior (higher Boundary F-score) and more coherent contours, especially in ambiguous border regions.
Open Zoom