An Interpretable Cost-Aware Framework for Mitigating Bias in Skin Lesion Classification Across Diverse Skin Tones

Cost-Aware EfficientNet (CAEN) achieves up to 87% recall in skin lesion classification across diverse skin tones, reducing bias in dermatological AI by 16.75% over previous models.

• Why it matters: Bias in skin lesion predictions limits equitable healthcare, especially for darker skin tones, due to underrepresentation and skewed datasets, risking misdiagnosis and health disparities.
• What they did: The study developed CAEN, a convolutional neural network with attention mechanisms and custom cost functions, optimized through iterative augmentation to improve fairness across skin tones.
• The result: CAEN demonstrated balanced, high recall rates and improved interpretability by focusing on relevant lesion regions, enabling fairer, more accurate dermatological AI for diverse populations.