Fairness aware subset selection for advancing equity in skin cancer detection

Area of Science:

• Artificial Intelligence in Medicine
• Dermatology
• Medical Diagnostics

Background:

• Skin cancer is a prevalent malignancy, with AI showing promise for early detection.
• Current AI models exhibit performance disparities, particularly underperforming on darker skin tones (Fitzpatrick Types V and VI).
• Existing fairness methods primarily focus on algorithmic adjustments, often overlooking crucial data quality and representation issues.

Purpose of the Study:

• To introduce FAIR-SCAN (Fairness and Accuracy through Ranking-Based Subset Selection for Skin Cancer Detection), a novel data-centric framework.
• To enhance fairness and accuracy in AI-driven skin cancer detection models.
• To address performance disparities across different skin tones in AI diagnostic tools.

Main Methods:

• FAIR-SCAN employs a data-centric framework utilizing subset selection guided by marginal contribution score (MCS) estimation.
• Data points are ranked based on their contribution to both model accuracy and fairness.
• An optimal subset of data is selected for training AI models, evaluated on Diverse Dermatology Images (DDI) and Fitzpatrick 17K datasets.

Main Results:

• FAIR-SCAN demonstrated improved balance in accuracy, True Positive Rate, and False Positive Rate across various skin tones.
• The method achieved this improvement while reducing the training dataset size by 50%.
• FAIR-SCAN outperformed existing algorithm-focused fairness methods in enhancing equity.

Conclusions:

• Strategic data selection is paramount for mitigating bias in AI-driven diagnostics.
• FAIR-SCAN's data-centric approach significantly enhances both the precision and equity of skin cancer detection.
• This framework supports the development of trustworthy and clinically deployable AI systems for skin cancer diagnosis.