Ano-Skin: Clinical Feature-Aware Diffusion Model for Dermatological Image Anonymization

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

Keywords: Dermatological Image Anonymization, Diffusion Model, Clinical Feature Preservation, Skin Disease Synthesis, Multi-modal prompting, Human Preference Opimization

TL;DR: "Ano-Skin" introduces a dermatological image anonymization method using diffusion models. It preserves critical diagnostic features, ensures disease-specific characterization, and enhances clinical utility through Simple Preference Optimization.

Abstract: Medical image anonymization requires effectively balancing patient privacy with clinical feature preservation, yet existing methods either compromise privacy or obscure critical disease information. We propose "Ano-Skin," a framework based on Stable Diffusion-v2 Inpainting with three key contributions: (1) Focused Feature Enhancement loss, $\boldsymbol{L_{FFE}}$ for preserving disease-specific characteristics, (2) Disease Difference loss, $\boldsymbol{L_{DIFF}}$ to maintain distinct visual patterns between skin conditions, and (3) Simple Preference Optimization ($\mathbf{SimPO}$) for seamless integration between preserved pathology and anonymized regions. Our method enables flexible control through mask and text-based prompting while maintaining high clinical utility across diverse skin conditions. Evaluated on 6,000 dermatological images, Ano-Skin significantly outperforms existing methods in disease classification performance (94.7% AUC retention), anonymization success (100%), and clinical assessment (4.5/5 by dermatologists). This work advances medical data sharing by resolving the traditional trade-off between privacy protection and diagnostic value in dermatological imaging.

Submission Number: 19