Skin Type Diversity in Skin Lesion Datasets: A Review.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Skin type diversity in image datasets is crucial for evaluating model performance across populations.
• Under-representation of darker skin types in datasets leads to decreased accuracy in deep learning models.
• Previous assessments have not fully evaluated the reporting and diversity of skin types in datasets.

Purpose of the Study:

• To investigate the general issue of skin type diversity in skin lesion datasets.
• To evaluate publicly available skin lesion datasets for reporting completeness and skin type representation.
• To identify shortcomings in current datasets and propose methods for verifiable diversity.

Main Methods:

• Systematic review of publicly available skin lesion datasets.
• Evaluation of dataset metadata for skin type reporting frequency and completeness.
• Analysis of skin type distribution and representation within the examined datasets.

Main Results:

• Identified significant variability in the reporting and assessment of skin type diversity across datasets.
• Found under-representation of certain skin types, particularly darker tones, in many skin lesion datasets.
• Highlighted the need for standardized reporting and improved data collection to ensure equitable AI model development.

Conclusions:

• Addressing skin type diversity in datasets is essential for developing robust and equitable AI in dermatology.
• Future efforts should focus on creating inclusive datasets with verifiable skin type representation.
• Improved data practices will enhance the generalizability and fairness of AI models across diverse populations.