Year: 2026
Type: article
Abstract: Melanoma is one of the deadliest forms of skin cancer, and early diagnosis depends on accurate delineation of suspicious lesions in dermoscopic images. This study introduces MaSA-UNet, a U-Net-based segmentation framework augmented with Manhattan Self-Attention (MaSA) for spatial context modeling, self-supervised denoising pretraining for representation learning, and a weighted composed loss (WCL) for overlap-aware optimization. The model is evaluated on four public skin-lesion benchmarks: ISIC 2016, ISIC 2017, ISIC 2018, and PH $$^2$$ 2 . In response to reviewer concerns, the revised study adds controlled multi-seed experiments, confidence intervals, paired statistical tests, isolated SSL comparisons, WCL sensitivity discussion, inference-time measurements, and a cross-dataset robustness discussion. The results show that MaSA-UNet remains competitive with strong U-Net baselines and can improve Dice and Jaccard on PH $$^2$$ 2 and ISIC 2016, while additional controlled experiments reveal dataset-dependent behavior on ISIC 2017 and ISIC 2018. These findings support MaSA as a useful contextual refinement mechanism for skin-lesion segmentation, while also clarifying the limitations of the current SSL and loss-weighting strategies. The source code will be made available in a GitHub repository upon publication.
Authors: Abel A. Reyes-Angulo, Sidike Paheding
Institutions: Michigan Technological University, Fairfield University
Language: English
Subfield: Oncology
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
Estimated APC paid: $3,390
OpenAlex ID: W7124919060
