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

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Artificial intelligence (AI) performance in dermatology is influenced by skin tone.
• Labeling datasets with skin tone data can enhance AI generalizability for detecting skin cancer.
• Large-scale validation of skin tone assessment methods for AI applications is needed.

Purpose of the Study:

• To assess the reliability of subjective (Fitzpatrick Skin Type [FST], Monk Skin Tone [MST], Pantone SkinTone Guide) and objective (colorimeter) tools for skin tone assessment.
• To evaluate the utility of these tools for labeling dermoscopic datasets in both in-person and photography-based settings.
• To determine the most effective method for capturing skin tone variations to improve AI performance in dermatology.

Main Methods:

• Prospective observational study comparing in-person and photography-based assessments.
• Utilized Fitzpatrick Skin Type (FST), Monk Skin Tone (MST), Pantone SkinTone Guide, and a colorimeter for skin tone evaluation.
• Correlated colorimeter measurements (gold standard) with subjective scale ratings and dermoscopic image color values.

Main Results:

• Colorimetry showed high precision for in-person measurements.
• Monk Skin Tone (MST) demonstrated high repeatability for both in-person and photography-based assessments, with better color space clustering than FST.
• Dermoscopic image color values poorly correlated with colorimetry; MST better reflected differences in AI melanoma classification scores compared to FST.

Conclusions:

• Fitzpatrick Skin Type (FST) is not a reliable proxy for skin tone in AI applications.
• The Monk Skin Tone (MST) scale is a reliable tool for assessing skin tone in dermatology AI.
• Accurate skin tone assessment is crucial for improving the performance and generalizability of AI algorithms in dermatology.