Removing facial features from structural MRI images biases visual quality assessment
Céline Provins1, Élodie Savary1, Thomas Sanchez1,2
1Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland.
Plos Biology
|April 30, 2025
Summary
Defacing human neuroimaging data to protect privacy alters human quality assessments but not automated MRIQC metrics. This impacts downstream analysis and interpretation of brain imaging data.
Area of Science:
- Neuroimaging
- Data privacy
- Image analysis
Background:
- Sharing human neuroimaging data requires removing facial features (defacing) for privacy.
- Defacing can remove non-identifiable information, potentially affecting data quality and analysis.
Purpose of the Study:
- To investigate how defacing impacts the quality assessment of T1-weighted human brain MRI scans.
- To compare the effects of defacing on both manual (human) and automated quality assessments.
Main Methods:
- Manual quality assessment by trained raters on defaced and non-defaced images from the IXI dataset (N=185).
- Automated quality assessment using MRIQC metrics on a larger IXI dataset subset (N=581) across three sites.
- Statistical modeling included linear mixed-effects models and repeated-measures, multivariate ANOVA (rm-MANOVA).
Main Results:
- Human raters' perception of image quality was significantly influenced by defacing.
- Automated quality metrics extracted by MRIQC were largely insensitive to the defacing process.
- Defacing introduces bias in manual quality assessment but not significantly in automated metrics.
Conclusions:
- Defacing, while necessary for privacy, introduces bias into manual quality assessments of neuroimaging data.
- Automated quality assessment tools like MRIQC show robustness against defacing-induced biases.
- Consideration of defacing's impact is crucial for reliable downstream analysis and interpretation of shared neuroimaging datasets.
