Assessment of Defacing Techniques on Medical Images for Radiation Therapy: Implications for Patient Privacy and Data Utility.

Abstract

To assess the impact of defacing-based deidentification techniques on reidentification risk and data utility across multimodal imaging in radiation therapy. We applied 4 defacing techniques: biometric_mask, quickshear, mri_reface, and Carina's deidentifier, to imaging from 88 brain patients (magnetic resonance imaging, computed tomography [CT], and RTDose) and 97 head and neck patients (positron emission tomography, CT, and RTDose) in The Cancer Imaging Archive. Reidentification risk was assessed using ArcFace, a deep learning-based facial recognition model, by measuring cosine similarity scores and conducting receiver operating characteristic analysis to distinguish between original and defaced images. Data integrity was evaluated by statistically comparing the volume and image intensity changes between the original and defaced images across 9 critical organs and gross tumor volume. Quickshear provides the highest privacy protection, achieving the lowest area under the curve across imaging modalities (area under the curve, 0.61-0.74), followed by Carina (0.59-0.80). Mri_reface showed moderate protection (0.70-0.91), whereas biometric_mask offered the least (0.76-0.94). Carina preserved structure volumes, and mri_reface produced minor volumetric changes in the eyes (5%) and lens (9%). In contrast, biometric_mask substantially affected the mandible (39%) and oral cavity (69%), whereas quickshear significantly altered multiple structures (10.0%-86.6%). Median changes in mean CT intensity after defacing were -48.3% (interquartile range [IQR], -65.9% to -33.9%) with biometric_mask and -77.7% (IQR, -89.4% to -51.7%) with quickshear in the oral cavity. For eyes, Carina and mri_reface produced changes of +24.5% (IQR, 11.6% to 38.2%) and +54.9% (IQR, 16.2% to 82.3%), respectively. In the brain data set, biometric_mask and quickshear decreased oral cavity D_mean by 0.72 Gy (IQR, 0.33 Gy to 0.89 Gy) and 1.61 Gy (IQR, 0.72 Gy to 2.12 Gy), respectively. Carina and mri_reface reduced eyes D_max by 2.11 Gy (IQR, 0.00 Gy to 3.39 Gy) and 1.05 Gy (IQR, 0.21 Gy to 1.16 Gy), respectively. A similar trend was observed in the head and neck data set with larger deviations. Carina's deidentifier and mri_reface showed favorable privacy-utility trade-offs relative to facial removal; the optimal choice may vary by application priorities.