Derm-FairAnon: Mitigating Demographic Bias in Skin Image Anonymization with Self-supervised Preference Optimization

YeonGyu Han, Jung Im Na, Seong Hwan Kim, Dongheon Lee

Keywords: Dermatological Image Anonymization, Diffusion Model, Clinical Feature Preservation, Skin Disease Synthesis, Bias Mitigation, Human Preference Opimization

TL;DR: We propose Derm-FairAnon, a novel framework that employs self-supervised preference optimization and distribution alignment to generate anonymized skin disease images with reduced demographic bias while preserving critical clinical features.

Abstract: Medical image anonymization faces the critical challenge of balancing patient privacy protection, clinical feature preservation, and demographic fairness. Existing methods often compromise privacy, obscure essential disease information, or perpetuate demographic biases in the anonymized outputs. We propose " Derm-FairAnon" a comprehensive framework for dermatological image anonymization that addresses these challenges through a novel integration of Stable Diffusion-v2 Inpainting with two key contributions: (1) Self-Supervised Preference Optimization ($\mathbf{selfPO}$), a novel approach that eliminates the need for explicit preference labels by leveraging image augmentation to generate self-supervised ranking signals; and (2) a demographic fairness mechanism with Skin-Fair loss, $\boldsymbol{L_{Skin-Fair}}$ that enables balanced demographic representation in generated images, effectively mitigating attribute biases while maintaining clinical utility. Evaluated on dermatological images from multiple hospitals, Skin-AnoFAIR outperforms existing methods in disease classification performance, anonymization success, demographic bias reduction, and clinical assessment by dermatologists.

Submission Number: 20