Abstract
Recent advances in magnetic resonance (MR) scanner quality and the rapidly improving nature of facial recognition software have necessitated the introduction of MR defacing algorithms to protect patient privacy. As a result, there are a number of MR defacing algorithms available to the neuroimaging community, with several appearing in just the last five years. While some qualities of these defacing algorithms, such as patient identifiability, have been explored in previous works, the potential impact of defacing on neuroimage processing has yet to be explored. We qualitatively evaluate eight MR defacing algorithms on 179 subjects from the OASIS-3 cohort and the 21 subjects from the Kirby-21 dataset. We also evaluate the effects of defacing on two neuroimaging pipelines—SLANT and FreeSurfer—by comparing the segmentation consistency between the original and defaced images. Results show that defacing can alter brain segmentation and even lead to catastrophic failures, which are more frequent with some algorithms such as Quickshear, MRI_Deface, and FSL_deface. Compared to FreeSurfer, SLANT is less affected by defacing. On outputs that pass the quality check, the effects of defacing are less pronounced than those of rescanning, as measured by the Dice similarity coefficient. We conclude that the effects of defacing are noticeable and should not be disregarded. Extra attention should be paid to the possibility of catastrophic failures. It is crucial to adopt a robust defacing algorithm and perform a thorough quality check before releasing defaced datasets.
Introduction
As MRI resolution improves and facial recognition AI becomes ubiquitous, "de-identifying" brain scans has moved beyond simply removing names from metadata. To prevent re-identification from 3D facial reconstructions, researchers use "defacing" algorithms to strip away facial features. However, defacing is not a "free lunch" for data integrity. Automated brain segmentation tools—the software used to measure brain volumes and cortical thickness—are highly sensitive to image changes. If a defacing algorithm accidentally clips a bit of the frontal lobe or alters the image's global intensity, the resulting scientific measurements might be compromised. This study evaluates the reproducibility of two major segmentation pipelines when applied to defaced data.
Methodology
The study performed a comprehensive evaluation of eight different defacing algorithms, ranging from simple geometric shearing to sophisticated deep-learning models and template-based blending. The algorithms evaluated include Defacer, Quickshear, Pydeface, MRI_Deface, FSL_deface, Face_Masking, AnonyMI, and mri_reface (versions 0.2 and 0.3).
The resulting images were processed through two prominent neuroimaging pipelines: FreeSurfer (a traditional, atlas-based pipeline) and SLANT (a modern deep-learning approach using 3D CNNs). The study utilized two public datasets: OASIS-3 (179 T1-weighted scans) for primary evaluation and Kirby-21 (21 subjects with scan-rescan data) to provide a real-world baseline for measurement error. Consistency was measured using the Dice Similarity Coefficient (DSC) between segmentations of unaltered and defaced versions of the same scan.
Results and Discussion
Catastrophic Failures and Pipeline Robustness
A primary finding was the prevalence of "Type II" failures—where defacing algorithms accidentally remove brain tissue. Several widely used tools, particularly Quickshear, MRI_Deface, and FSL_deface, frequently committed these errors. For instance, Quickshear failed on 101 out of 179 OASIS-3 subjects.
Regarding the segmentation pipelines, SLANT (the deep-learning approach) proved significantly more robust than FreeSurfer. SLANT achieved a high mean DSC of 0.970, whereas FreeSurfer was lower and more variable at 0.918, showing a wider spread of errors and dramatic outliers even when brain tissue remained physically intact.
The "Scan-Rescan" Benchmark
Crucially, on images that passed manual quality checks, the variation introduced by defacing was actually smaller than the natural variation found in scan-rescan data. This suggests that with proper oversight, defacing is a scientifically acceptable practice that does not introduce more noise than standard longitudinal scanning variability.
Action at a Distance
Interestingly, defacing can cause segmentation errors in regions far from the face, such as the occipital lobe. This occurs because removing facial voxels changes how the software registers or aligns the brain to a standard template, demonstrating that the impact of defacing is global rather than just local to the face.
Conclusion
The study recommends mri_reface (version 0.3) as the top-performing algorithm, as it produced zero catastrophic failures and maintained high segmentation consistency. While defacing is essential for privacy, it necessitates manual quality checks to avoid catastrophic data loss. To improve reliability, researchers are encouraged to include multiple brain segmentation pipelines, particularly deep-learning models like SLANT, which demonstrate greater resilience to the image modifications introduced by defacing.
Summary updated