Reproduce “Colorimeter-Supervised Skin Tone Estimation from Dermatoscopic Images for Fairness Auditing” on OpenResearch

About this paper

Marin Benčević, Krešimir Romić, Ivana Hartmann Tolić, Irena Galić

Researchers from Josip Juraj Strossmayer University of Osijek developed a neural network framework to accurately estimate Fitzpatrick skin types and Individual Typology Angle (ITA) from dermatoscopic images, supervised by gold-standard colorimeter measurements and expert annotations. The framework achieved an Intraclass Correlation (ICC3) of 93.88% for ITA and a linear-weighted Cohen's kappa of 52.98% for Fitzpatrick types, subsequently revealing that widely used public datasets are severely imbalanced, with less than 1% representation of Fitzpatrick types V and VI.

Key insights

• Neural networks can reliably learn the complex, non-linear relationship between dermatoscopic image pixels and objective physical skin tone measurements (colorimeter-derived CIELAB/ITA).
• Ordinal regression proves more effective for predicting categorical Fitzpatrick types than standard classification, yielding more reliable and accurate results.
• The majority of publicly available dermatoscopic datasets are heavily biased towards lighter skin tones, with less than 1% representation of Fitzpatrick types V and VI.

Summary by alphaXiv.