Leveraging Atlas-Based Segmentation for Anatomy-Preserving Facial De-Identification in Head CTs to Protect Patient Privacy

Introduction: Modern facial recognition software has introduced the risk of patient identification from three-dimensionally reconstructed computed tomography (CT) scans. This has prompted the need for reliable facial de-identification methods to protect patient privacy. Current facial de-identification methods often obscure or destroy sinuses, turbinates, and other relevant anatomy in otolaryngology. We present a facial de-identification method that preserves these structures, as well as an automated workflow for rapidly de-identifying large-volume datasets.

Methods: A total of 21 head CTs were obtained from the New Mexico Decedent Image Database (NMDID). Scans of patients with oral or nasal lines/tubes/airways were not included in this study. Using 3D Slicer, a seed-growing segmentation technique was performed to label the skin around the face. This label was dilated bidirectionally to form a 6-millimeter mask to obscure facial features. To automatically de-identify CTs, an atlas-based segmentation method was developed, in which a single head CT and corresponding de-identification mask was deformed to match the contours of other head CTs. Accuracy of this automated method against manually generated masks were then evaluated with Dice scores.

Results: Manual de-identification of the face was successful, preserving nasal turbinates and sinuses in all included datasets upon inspection by two independent reviewers. De-identification masks generated by automated atlas-based segmentation overlapped manually generated face masks with a Dice score of 0.71±0.09. This automated facial de-identification algorithm completed with an average runtime of 157.2±19.1 seconds per scan.

Conclusions: We have demonstrated a reliable facial de-identification method for head CTs, as well as an automated form of this method for batch de-identifying image datasets. These techniques show promise for preventing patient identification while preserving underlying sinonasal anatomy for medical image analysis.