Answers are grounded in this paper and cite their sources.
Deep learning models for skin lesion segmentation and classification achieve strong aggregate accuracy on dermoscopic benchmarks but generalise poorly across skin tones owing to the underrepresentation of darker Fitzpatrick types in public training datasets. Existing fairness-aware methods address this disparity exclusively at the classification stage, implicitly assuming that upstream segmen-tation is unbiased; an assumption increasingly contradicted by the literature. We propose FSD-Net (Fairness-aware Skin Disease Network), which to our knowledge is the first framework to enforce and evaluate fairness simultaneously at both the 1 segmentation and classification stages. FSD-Net couples a ViT-B/16 segmenta-tion backbone conditioned on a continuous skin tone score through Feature-wise Linear Modulation with a Swin-V2-Tiny classifier operating on segmentation-guided region-of-interest crops. Evaluated on ISIC-2017, ISIC-2018, PH2, and the patient-disjoint HAM10000 benchmark, FSD-Net achieves Dice coefficients of 0.851, 0.892, and 0.927 on the three segmentation datasets and 84.2% classification accuracy with an Equalized Opportunity gap of 0.025, reducing to 0.004 with test-time augmentation. A six-configuration ablation study attributes a 3.68 percentage-point Dice improvement on ISIC-2017 directly to the FiLM module. Bootstrap confidence intervals and McNemar pairwise tests confirm that fairness improvements are statistically reliable.
No takes yet. Share an insight, caveat, or question.
Khan et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: