Author: Ali Ammar, Quan Chen, Yi Rong, Libing Zhu 👨🔬
Affiliation: Mayo Clinic Arizona 🌍
Purpose: To investigate how defacing algorithms, essential for patient privacy in data sharing, impact AI-based segmentation performance in CT imaging for radiation therapy. This study evaluates whether anonymization affects segmentation accuracy, particularly for non-facial structures critical to clinical and research applications.
Methods: This study evaluated the impact of defacing algorithms on segmentation performance using head-and-neck (HN) CT data from 50 patients (mean age 63.5±19.9 years; 24F, 26M). Two defacing algorithms, DeIdentifier (Carina Medical LLC) and mri_reface 0.3.3 (open-source), were applied to anonymize images by obscuring front facial features while preserving anatomical integrity. Segmentation was performed using two auto-segmentation platforms, INTContour (Carina Medical LLC) and AccuContour® (Manteia Technologies Inc). The experiments compared (1) both defacing algorithms using INTContour and (2) INTContour versus AccuContour on CT images defaced with DeIdentifier. Segmentation accuracy was evaluated using Dice coefficients, Hausdorff Distance at the 95th-percentile (HD95), and Spatial Dice Similarity Coefficient (SDSC) with a 2 mm tolerance.
Results: DeIdentifier and mri_reface showed distinct impacts on segmentation performance. DeIdentifier maintained stable accuracy across facial and deeper structures, with Dice coefficients (0.85–0.9), low HD95 values (1.5 mm), and consistent surface alignment (SDSC_2mm near 1.0). In contrast, mri_reface exhibited greater variability, with Dice scores (0.7–0.85) and higher HD95 values, particularly in facial regions and small non-facial structures like the optic nerve and parotid gland. INTContour and AccuContour were similarly affected by defacing, with INTContour achieving stability in most regions but showing variability in smaller structures (e.g., cochlea, brachial plexus). AccuContour remained consistent in deeper structures but exhibited variability in facial regions and smaller structures.
Conclusion: Defacing algorithms affect segmentation performance in both facial and non-facial structures. While effective, their impacts depend on anatomical location and structure complexity. These findings emphasize the need to validate tools to ensure anonymized datasets remain reliable and usable.