De-identification Strategy and Re-identification Risks for Facial Computed Tomography Images via Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Head computed tomography (CT) scans contain sensitive facial information, posing re-identification risks.
• Existing de-identification methods may compromise facial bone structure integrity.
• Effective privacy protection is crucial for sharing medical imaging data.

Purpose of the Study:

• To develop and evaluate a deep learning-based selective de-identification method for head CT images.
• To remove facial soft-tissue features while preserving facial bone structures.
• To assess re-identification risk after de-identification for robust privacy protection.

Main Methods:

• A retrospective study included 3206 facial CT scans from 3091 patients.
• A YOLOv8-based model was used to selectively remove facial soft-tissue features.
• Re-identification risk was assessed using deep learning facial embeddings and human evaluation.

Main Results:

• The de-identification model achieved a mean average precision (mAP) of 0.858.
• Deep learning-based re-identification accuracy decreased from 85% to 64%.
• Human re-identification accuracy dropped from 84% to 55% for both general participants and plastic surgeons.

Conclusions:

• A selective de-identification method for facial CT images was successfully developed.
• The method effectively preserves craniofacial structures and significantly reduces re-identification risk.
• The publicly available model and demo facilitate broader adoption in facial CT research.