Mitigating Skin Tone Bias in Lesion Screening via Confidence-Based Routing
International Journal of Scientific Research in Engineering and Management · 2026-10 · 10 (10)
DOI: 10.55041/ijsrem68502 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Scientific Research in Engineering and Management, 2026-10. Verfasst von Khasim Syed und 4 weiteren Autor:innen.
Abstract (Originalsprache)
Deep-learning systems for skin-lesion screening are usually reported with one aggregate accuracy on one dataset, which says little about how they behave across skin tones, on images from other sources, or on photographs that are not lesions at all. We built DermaScan AI, a web-based pipeline that combines a DermNet-trained ResNet18 pre-filter, a three-tier (benign, monitor, urgent) ResNet18 risk classifier, Grad-CAM ex-planations, and a confidence gate whose threshold depends on the user’s Fitzpatrick skin type. Our questions were how a baseline classifier behaves across skin types, whether a changed training recipe helps, and how far accuracy falls on data from another source. A baseline trained on HAM10000 reached 80.84% vali-dation accuracy but only 44.2–51.9% on held-out Fitzpatrick17k images, depending on skin type. Retraining on HAM10000 plus half of the lesion-relevant Fitzpatrick17k images, with a lower learning rate, a scheduler and rotation augmentation (v3), raised validation accuracy to 88.69% and held-out accuracy to 54.4–74.8%. The gains for types I–V (14.2–25.5 points) exceeded run-to-run variation over three seeds (SD 1.1–2.2 points); the Type VI gain (2.5 points) did not. The gap between the best and worst skin type widened from 7.7 to 20.3 points. On ISIC 2019 images that are not in HAM10000 (n = 2, 403), balanced accuracy was 55.70% against 80.50% on the HAM10000 validation split. The pre-filter reached 86.20% in-domain. Splits are image-level, labels are tiers we derived from dataset diagnoses, no clinician reviewed any output, and the routing rule was not evaluated by risk–coverage analysis, so these results say nothing about clinical safety.
Zitieren
APA
Syed, K., Kumar, M. M., Jayadithya, A., Varshitha, L., & Sri, M. B. (2026). Mitigating Skin Tone Bias in Lesion Screening via Confidence-Based Routing. International Journal of Scientific Research in Engineering and Management, 10(10). https://doi.org/10.55041/ijsrem68502
Vancouver
Syed K, Kumar MM, Jayadithya A, Varshitha L, Sri MB. Mitigating Skin Tone Bias in Lesion Screening via Confidence-Based Routing. International Journal of Scientific Research in Engineering and Management. 2026;10(10). doi:10.55041/ijsrem68502
BibTeX
@article{Syed2026,
author = {Syed, Khasim and Kumar, M. Mukesh and Jayadithya, A. and Varshitha, L. and Sri, M. Bhagya},
title = {Mitigating Skin Tone Bias in Lesion Screening via Confidence-Based Routing},
journal = {International Journal of Scientific Research in Engineering and Management},
year = {2026},
volume = {10},
number = {10},
doi = {10.55041/ijsrem68502}
}
EUVIMED