Evaluating skin tone scales for dermatologic dataset labeling: a prospective-comparative study

Skin tone affects artificial intelligence (AI) performance in dermatology. While labeling datasets for skin tone could improve algorithm generalizability for detecting dermatologic malignancies, large-scale validation of skin tone assessments is lacking. This prospective observational study assessed reliability of subjective tools (Fitzpatrick Skin Type [FST], Monk Skin Tone [MST], Pantone SkinTone Guide) and an objective colorimeter for in-person and photography-based settings to evaluate utility for labeling dermoscopic datasets. Colorimetry (gold standard for color measurement) demonstrated high precision with in-person measurements. Of subjective scales, MST demonstrated slightly tighter clustering in the color space and high repeatability for in-person and photography-based assessments (latter varied by lighting). Dermoscopic image-extracted color values correlated poorly with colorimetry values. For subjective ratings, MST more effectively captured differences in AI melanoma classification scores than FST. Findings underscore that FST is not a proxy for skin tone; an important role remains for skin tone assessment to improve AI performance.