De-identification and Obfuscation of Gender Attributes from Retinal Scans

De-identification and Obfuscation of Gender Attributes from Retinal Scans

연구 분야: Analysis

학회: Workshop on Clinical Image-Based Procedures, Workshop on the Ethical and Philosophical Issues in Medical Imaging, MICCAI Workshop on Fairness of AI in Medical Imaging

Retina images are considered to be important biomarkers and have been used as clinical diagnostic tools to detect multiple diseases. We examine multiple techniques for de-identifying retina images while maintaining their clinical ability for detecting diabetic retinopathy (DR), using gender as a proxy for identifiability. We apply two differential privacy algorithms, Snow and VS-Snow, on the entire image (globally) and on blood vessels only (locally) to obfuscate important image features that can predict a patient’s sex. We evaluate the level of privacy and retained clinical predictive power of these de-identified images by using attacking gender classifier models and downstream disease classifiers. We show empirically that our proposed VS-Snow framework achieves strong privacy while preserving a meaningful clinical predictive power across different patient populations.