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

TL;DR

The paper tackles privacy risks from facial features in routine head MRIs and the impact of de-identification on brain morphometry. It introduces a novel 3D conditional GAN-based refacing method trained to replace defaced faces with anonymized but structure-preserving faces. The authors compare this approach against defacing and other refacing tools across brain volumetry reproducibility, re-identification risk, and processing time, using two segmentation tools and ArcFace for risk estimation. They find that the 3D cGAN offers strong reproducibility and competitive privacy protection, with rapid inference, particularly when applied to defaced data, highlighting a practical privacy–post-processing fidelity trade-off in neuroimaging data sharing.

Abstract

With the rise of open data, identifiability of individuals based on 3D renderings obtained from routine structural magnetic resonance imaging (MRI) scans of the head has become a growing privacy concern. To protect subject privacy, several algorithms have been developed to de-identify imaging data using blurring, defacing or refacing. Completely removing facial structures provides the best re-identification protection but can significantly impact post-processing steps, like brain morphometry. As an alternative, refacing methods that replace individual facial structures with generic templates have a lower effect on the geometry and intensity distribution of original scans, and are able to provide more consistent post-processing results by the price of higher re-identification risk and computational complexity. In the current study, we propose a novel method for anonymised face generation for defaced 3D T1-weighted scans based on a 3D conditional generative adversarial network. To evaluate the performance of the proposed de-identification tool, a comparative study was conducted between several existing defacing and refacing tools, with two different segmentation algorithms (FAST and Morphobox). The aim was to evaluate (i) impact on brain morphometry reproducibility, (ii) re-identification risk, (iii) balance between (i) and (ii), and (iv) the processing time. The proposed method takes 9 seconds for face generation and is suitable for recovering consistent post-processing results after defacing.

Figures (16)

• Figure 1: Summary of experiments conducted within the current study. In addition to the presented experiments the processing time of each de-identification tool was evaluated.
• Figure 2: Overview of the proposed 3D cGAN including generator and discriminator architectures. 4D objects are visualized via 3D projections by collapsing the depth axis, hence only Height (H), width (W) and channels (C) axes are displayed.
• Figure 3: Examples of the de/refacing techniques on one subject
• Figure 4: Bland-Altman difference plots for the volumetric results of the original images in comparison to the ones of the de/refaced images. Plots one big region (TIV) and the hippocampus as pars pro toto for a small region.. Note that vertical axes scaling differs for fsl_anat and MorphoBox.
• Figure 5: Trade-off plots for the joint evaluation of the re-identification risk and reproducibility of the morphometry results after de/refacing. The inverse face distance averaged across subjects is plotted on the y-axis as a measure of the re-identification risk. The vertical whiskers' length is the standard deviation of the inverse distances. The normalized coefficients of repeatability (nCR) were calculated using the normalized estimated volumes (averaging both across scans and across brain structures) are displayed on the x-axis as a measure of the inconsistency in the volumetric results after face de-identification. Horizontal whiskers have the length of the standard deviation of the CR values, calculated across scans for different normalized brain structure volumes.