Abstract
Dermatology AI models are trained predominantly on light-skinned, cancer-focused image collections, yet are increasingly proposed for deployment in resource-constrained settings (RCS) whose patients differ from the training population along two confounded axes at once: skin tone and disease distribution. When such models underperform, the field has largely attributed the failure to skin-tone underrepresentation. Using only public datasets and free-tier compute, we disentangle these factors. We evaluate a cancer-trained baseline (ResNet-50 finetuned on HAM10000 and ISIC 2019), two open dermatology foundation models (DermLIP, MONET), and a general-purpose vision model (DINOv3) as frozen feature extractors, probing them on a tone-varying but disease-matched dataset (DDI) and a disease-shifted, tone-diverse dataset (SCIN). In our setting, disease-distribution shift contributes more than skin tone: the cancer baseline degrades sharply from its in-domain performance to unfamiliar conditions (a drop from 0.62 to 0.21 balanced accuracy across clinical categories), a far larger effect than the within-disease skin-tone gap, which is comparatively small (0.10-0.18) and did not show a consistent direction. Using a label-free representation-quality analysis, we show this collapse is representational, not merely an artifact of disjoint label spaces: cancer-specialized features barely cluster unfamiliar conditions (kNN neighbor-purity lift +0.06 over chance), whereas dermatology-pretrained features retain substantially more transferable structure (+0.23). Finally, latent representation quality closely tracks recoverable performance under cheap, no-retraining adaptation, suggesting a promising direction for RCS deployment: beginning from a dermatology foundation model, where roughly ten labeled examples per condition recover most of the attainable performance. We release code and the evaluation protocol to support reproducible auditing.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_077.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=oUPHj1WDAU
BibTex
@InProceedings{KunNir_Disease_MICCAISAT2026,
author = { Kunwor, Nirajan AND Poudel, Sanjaya AND Trinh, Quoc-Huy AND Arafat, Jahidul AND Gaire, Sunil Kumar},
title = { { Disease Burden over Skin Tone: Decomposing the Dermatology-AI Generalization Gap } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17265},
month = {pending},
page = {pending}
}
