1156 - NIMOD: A Noninvasive Multimodality Oncology Defacer with Near-Chance Reidentification Risk and Preserved Data Utility

Purpose/Objective(s): Defacing medical images is essential for privacy-compliant data sharing in radiation oncology, yet existing techniques force a tradeoff between privacy and data utility. Prior evaluation showed that the most aggressive face-region removal caused up to 87% volume loss in critical structures such as the mandible and oral cavity, while anatomy-preserving methods still altered voxel intensities by 24–44% for superficial OARs. This study presents NIMOD, an anatomy-guided defacer that reduces the need for tradeoff significantly by achieving near-chance reidentification risk while preserving exact voxel intensities within the volumes of interest (VOIs) across sites and modalities.

Materials/Methods: NIMOD generates a face mask from TotalSegmentator segmentation (body contour and skeletal structures), defining facial soft tissue while excluding brain, bone, and all VOIs. Facial voxels are removed via modality-specific methods: replacement with air (background HU) using per-slice ellipsoid fitting and Gaussian blending for CT, Laplace diffusion inpainting from surrounding tissue for dose, gradient interpolation from boundary SUV toward background for PET, and texture-matched synthesis for MRI. Original values within all VOIs are exactly restored. We evaluated 185 cases across two cohorts: 88 brain tumor cases and 97 head and neck (HN) cancer cases from TCIA, spanning CT, MRI, PET, and dose modalities. The HN cohort includes large field-of-view scans with variable head positioning and arms-up acquisition. Data utility was verified by confirming zero voxel difference within 14 VOIs for CT and dose. Reidentification risk for all modalities was evaluated using ArcFace on 2D frontal snapshots from 3D surface renderings, computing cosine similarity across all N × N original-to-defaced pairs (N = subjects per cohort) with ROC analysis.

Results: For CT and dose across all 185 cases, every voxel within every VOI was preserved with zero difference, confirming exact data preservation. NIMOD achieved near-chance reidentification AUC across all cohort–modality combinations: brain CT 0.586, MRI 0.576, dose 0.554; HN CT 0.521, PET 0.517, dose 0.571. These results surpass conventional full-face defacing (AUC 0.65–0.74), while NIMOD preserves all voxel values within VOIs that such methods destroy. The segmentation-defined boundaries generalized across institutions and modalities, including challenging HN cases where critical OARs (mandible, oral cavity, parotid glands) lie adjacent to the defacing boundary.

Conclusion: NIMOD is one of the first defacing solutions designed specifically for radiation oncology and validated across multiple modalities (CT, MRI, PET, dose) as well as anatomical regions (brain, HN). It achieves lower reidentification risk than conventional full-face defacing while maintaining exact voxel fidelity within all VOIs for CT and dose, enabling safe multimodal data sharing for multicenter research without compromising downstream applications.