Analyzing the TotalSegmentator for facial feature removal in head CT scans

Area of Science:

• Medical Imaging
• Computer Vision
• Data Privacy

Background:

• Facial recognition technology poses privacy risks in medical imaging, especially with head CT scans containing identifiable facial features.
• Protecting patient data while enabling research is a critical challenge in healthcare.

Purpose of the Study:

• To evaluate the efficacy of facial recognition software in identifying facial features from head CT scans.
• To explore a defacing pipeline using TotalSegmentator for reducing re-identification risks.
• To assess the impact of defacing on data integrity for research purposes.

Main Methods:

• Analysis of 1404 head CT renderings (defaced and non-defaced) from the UCLH EIT Stroke dataset.
• Comparison of TotalSegmentator's defacing performance against a state-of-the-art CT defacing algorithm.
• Utilized deep learning for face detection and compared cosine similarity of facial embeddings using a Support Vector Machine classifier with 5-fold cross-validation.

Main Results:

• Facial features were detected in 76.5% of non-defaced CT scans.
• Defacing with TotalSegmentator significantly reduced re-identification performance (ROC-AUC 0.55, accuracy 0.56) compared to non-defaced images (ROC-AUC 0.69, accuracy 0.65).
• TotalSegmentator offered slightly superior privacy protection over the CTA-DEFACE algorithm.

Conclusions:

• Facial recognition software can identify individuals from head CT scans.
• The TotalSegmentator defacing pipeline effectively minimizes re-identification risks to near-chance levels.
• This privacy-preserving pipeline is valuable for multi-site research, data sharing, and compliance with medico-legal requirements.