An Interpretable Cost-Aware Framework for Mitigating Bias in Skin Lesion Classification Across Diverse Skin Tones
Journal:
medRxiv
Published Date:
Sep 2, 2026
Abstract
Despite advances in dermatological AI, skin lesion predictions continue to exhibit significant bias, consistently exhibiting underperformance in brown and darker tones. This disparity stems largely from the lack of representation in commonly used datasets such as Fitzpatrick17k and Diverse Dermatology Images (DDI), which are heavily skewed toward lighter skin tones. Existing models, including those trained on balanced datasets like Skin Cancer Benign vs. Malignant, show steep drops in recall when evaluated across varying pigmentation. To address this inequity, the Cost-Aware EfficientNet (CAEN) model based on EfficientNet-B0, a widely adopted convolutional neural network, is proposed, incorporating attention mechanisms and custom cost functions to better capture underrepresented tones and reduce class imbalance. Incorporating an optimization-based augmentation strategy further enhances fairness by iteratively determining the optimal augmentation ratio for each skin tone, resulting in a more balanced training. CAEN achieved average-recall rates of 86% for non-neoplastic, 87% for benign, and 87% for malignant lesions in three datasets, with a performance gain of 16.75% over prior studies. The model also demonstrated robustness under artificial test conditions, including variations in brightness and contrast, with marked improvement for brown and darker skin tones. Interpretability analysis using Gradient-Weighted Class Activation Mapping (Grad-CAM) showed that baseline EfficientNet-B0 often focuses on irrelevant image regions in darker skin tones. CAEN corrected these issues by attending more accurately to areas relevant to the lesion, improving both transparency and equity. This approach marks a step toward fair and generalizable lesion prediction, addressing long-standing performance disparities across diverse pigmentations in dermatological AI.
