Image-derived skin-tone estimates do not recover clinician-recorded Fitzpatrick type in smartphone dermatology photographs
Year: 2026
Type: preprint
Abstract: Background. Evaluations of dermatology artificial-intelligence (AI) tools increasingly report performance by skin tone. When skin-tone labels are missing, a common shortcut is to estimate tone from the image itself, usually with the individual typology angle (ITA). Whether such estimates agree with clinician skin typing on everyday clinical photographs has rarely been tested directly. Methods. We applied an ITA-based estimator (median ITA of skin-colored pixels in a border frame around the centered lesion, mapped to six classes with conventional cut-points) to all 1,494 smartphone lesion photographs in the public PAD-UFES-20 dataset that carry a clinician-recorded Fitzpatrick type. We report exact and within-one-class agreement, quadratic-weighted kappa and Spearman correlation, with 95% confidence intervals from a patient-clustered bootstrap, and compare them with chance (label permutation) and with a constant guess. Results. Of 1,494 images, 1,492 (from 794 patients) yielded an estimate. Exact agreement was 18.4% (95% CI 16.5-20.5) against 17.7% expected by chance; agreement within one class was 54.1% (51.2-56.9) against 54.5% by chance. Quadratic-weighted kappa was -0.02 (-0.06 to 0.02) and Spearman rho -0.02 (-0.08 to 0.04). Always guessing type II, the most common type, would have been exactly right for 58.6% of images. Median ITA was nearly identical across clinician types I to IV (31 to 36 degrees). Results were unchanged when restricted to images the estimator itself rated as good quality, when the whole image was used, and on full-resolution originals. Conclusions. On smartphone clinical photographs taken under ordinary conditions, image-derived ITA carried no measurable information about clinician-recorded Fitzpatrick type. Subgroup evaluations of dermatology AI should not stratify by image-derived tone estimates of this kind; they should use clinician- or patient-assigned labels whose reliability is measured and reported.
Author: Pennie, Niya D.
Language: English
Subfield: Oncology
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7225636159