A Fairness-Aware Framework for Bias Mitigation in Diagnostic Skin Imaging Using an Adversarial CNN Approach

Abstract

The integration of Artificial Intelligence (AI) into dermatological diagnostics offers great potential for improving clinical decision- making and patient outcomes. However, persistent biases in training data and model structures continue to increase disparities in diagnostic accuracy, especially across different skin tones and lesion types. These inequalities emphasize the urgent need for fairness - aware methods that provide equitable performance in clinical settings. Therefore, this study aimed to develop and validate a fairness - aware AI framework for dermatological imaging, designed to reduce bias while maintaining diagnostic accuracy. The framework was guided by three goals: (i) to assess how well adversarial debiasing techniques can remove bias- related features, (ii) to analyze current diagnostic systems and identify fairness gaps, and (iii) to introduce a customized bias mitigation strategy for skin lesion classification. Adversarial debiasing was applied through a Gradient Reversal Layer (GRL) within a Convolutional Neural Network (CNN). The GRL works by reversing gradient signals during training, discouraging the network from encoding bias- related attributes while still optimizing for diagnosis accuracy. This mechanism formed the core of the fairness - aware framework. Performance was evaluated by comparing the framework to a baseline CNN using the ISIC 2020 skin lesion dataset, which includes 33, 126 dermoscopic images from over 2, 000 patients. Preprocessing and augmentation techniques were used to improve data quality and model robustness. External validation utilized the Fitzpatrick 17 k dataset, containing 16, 577 clinical images annotated by skin type, to test demographic generalizability. Expert validation was also performed, with dermatologists reviewing interpretability outputs to confirm clinical relevance and practical use. Model interpretability was improved through SHAP- based feature attribution, helping clinicians visualize decision boundaries and better understand the framework's diagnostic reasoning. The fairness- aware framework significantly reduced the Statistical Parity Difference from. .35 to. .05, while maintaining high diagnostic accuracy, sensitivity, and AUC--ROC scores above. .87. Fairness improvements were supported by Equalized Odds metrics, and statistical testing showed that these equity gains did not compromise reliability. These results highlight the clinical potential of fairness- aware AI in dermatology. By combining adversarial debiasing, expert validation, and practical application, the proposed framework offers actionable paths for the ethical use of AI in precision medicine. More broadly, it advances the discussion on bias mitigation in healthcare AI, demonstrating the transformative potential of equity- focused approaches to improve diagnosis outcomes across diverse populations.