Multisite Comparison of MRI Defacing Software Across Multiple Cohorts.

Area of Science:

• Neuroimaging
• Data Privacy
• Medical Informatics

Background:

• Advancements in neuroimaging and facial recognition increase risks of participant re-identification from structural scans.
• Removing facial features is essential for data sharing and open release of neuroimaging data.
• Existing defacing software lacks comprehensive accuracy reviews across diverse populations.

Purpose of the Study:

• To evaluate the accuracy of publicly available facial feature removal (deface) algorithms for structural neuroimaging scans.
• To compare the performance of six defacing algorithms and one skull stripping tool across varied age groups and patient cohorts.

Main Methods:

• Tested six defacing algorithms (afni_refacer, deepdefacer, mri_deface, mridefacer, pydeface, quickshear) and FreeSurfer on 300 structural MRI scans.
• Assessed accuracy based on complete facial feature removal and preservation of brain tissue.
• Validated defaced scans through preprocessing pipelines to check for downstream effects.

Main Results:

• Success rates varied significantly among algorithms; afni_refacer (89%) and pydeface (83%) performed best overall.
• Algorithm performance was influenced by dataset characteristics, with specific age groups posing challenges for each top-performing defacer.
• Defacing did not significantly alter downstream preprocessing results compared to original scans.

Conclusions:

• Afni_refacer and pydeface are the most effective publicly available tools for defacing neuroimaging data.
• Algorithm choice and performance depend on the specific neuroimaging dataset, particularly age demographics.
• Defacing neuroimaging data is a viable method for enhancing participant privacy without compromising data utility for future analyses.