Fast refacing of MR images with a generative neural network lowers re-identification risk and preserves volumetric consistency

Nataliia Molchanova ^1,2,3,4,5, Bénédicte Maréchal ^1,4,5, Jean-Philippe Thiran ^1,5

⁰Department of Radiology, Lausanne University Hospital (CHUV), Lausanne, Switzerland.

Human Brain Mapping +

|

June 20, 2024

View abstract on PubMed

Summary

A new 3D generative adversarial network method anonymizes 3D MRI head scans, balancing privacy protection with brain morphometry reproducibility. This approach offers faster processing and consistent results for open data initiatives.

Area Of Science

Medical Imaging
Computer Vision
Data Privacy

Background

3D renderings from structural MRI scans raise privacy concerns due to individual identifiability.
Existing de-identification methods like blurring, defacing, and refacing have limitations.
Defacing offers strong privacy but impacts post-processing; refacing preserves data but increases re-identification risk.

Purpose Of The Study

To propose and evaluate a novel method for anonymized face generation in defaced 3D T1-weighted MRI scans.
To assess the proposed method's performance against existing de-identification tools.
To analyze the trade-offs between re-identification risk, brain morphometry reproducibility, and processing time.

Main Methods

Development of a 3D conditional generative adversarial network for anonymized face generation.
Comparative study evaluating the proposed method against existing defacing and refacing tools.
Assessment using two segmentation algorithms (FAST and Morphobox) to analyze impact on brain morphometry, re-identification risk, and processing time.

Main Results

The proposed method achieves anonymized face generation in 9 seconds.
It demonstrates suitability for recovering consistent post-processing results after defacing.
Comparative analysis evaluated impact on brain morphometry reproducibility, re-identification risk, and processing efficiency.

Conclusions

The novel 3D conditional generative adversarial network offers an effective solution for de-identifying 3D MRI head scans.
The method balances privacy protection with the need for consistent post-processing, particularly for brain morphometry.
This approach provides a promising tool for enhancing privacy in open data initiatives using neuroimaging data.

Keywords:

brain morphometry conditional generative adversarial networks defacing de‐identification magnetic resonance imaging privacy re‐identification risk