5. Defacing Workflow

5.1. Introduction

Defacing MRI data is crucial for protecting participant privacy in neuroimaging research. Structural brain MRI include facial features that can be reconstructed and used to identify individuals. Defacing removes these features while preserving brain anatomy for analysis.

5.2. Requirements

CPU	RAM
1	5 GB

5.3. Description

Processing Steps

• Defacing T1w image We use the defacing method that is provided as part of the standard FSL distribution under the command fsl_deface [1]. Similar to other established defacing tools, such as mri_deface [2] and pydeface [3], the method relies on linear registration to generate a mask that identifies facial voxels. Voxels within this mask are then set to zero to remove identifiable facial features. A key distinction of fsl_deface compared with mri_deface and pydeface is that it additionally removes the ears, providing a more comprehensive anonymization of head anatomy. This workflow is applied to T1-weighted structural images.

• Defacing T2w and FLAIR images The defacing mask obtained from the T1w image is propagated to other modalities through rigid alignement.

Quality Control

• Correlation score For each aligned image, we compute its correlation with the MNI for T1w or to T1w for T2w and FLAIR reference image. Images are then sorted in ascending order of this score, allowing potential outliers to be easily identified.

• Overlap score For each image, we compute the overlap ratio between the brain mask extracted using FreeSurfer’s deep‑learning–based mri_synthstrip method [4] and the corresponding defacing mask. Images are then sorted in ascending order of this score, allowing potential outliers to be easily identified.

• Manual inspection Following the overlap-based ranking, a manual quality control step is performed using the generated defacemosaic figure. This figure allows visual inspection of the defaced image and verification that facial/ear structures have been successfully removed. The most obvious outliers are thus removed.

• Thresholding The correlation score is thresholded at 0.5, meaning that if an image is not roughly registered to the template, the preprocessing is considered invalid. The overlap score is thresholded at 5%, meaning that if the defacing mask removes a substantial portion of the brain, the preprocessing is considered invalid. Images with a correlation lower than 0.5 or an overlap greater than 5% are flagged as low‑quality.

5.4. Outputs

The defacing directory contains subject-level results, logs, and quality-control outputs. The structure is organized following the brainprep ontology.

defacing/
├── dataset_description.json
├── figures
│   ├── histogram_mean_correlation_<T1w|T2w|FLAIR>.png
│   └── histogram_overlap_<T1w|T2w|FLAIR.png
├── log
│   ├── report_<timestamp>.rst
│   └── commands_<timestamp>.rst
├── quality_check
│   ├── mask_overlap_<T1w|T2w|FLAIR>.tsv
│   └── mean_correlations_<T1w|T2w|FLAIR>.tsv
└── subjects
    └── sub-01
        └── ses-01
            ├── figures
            │   └── sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>_defacemosaic.png
            ├── log
            │   ├── report_<timestamp>.rst
            │   └── commands_<timestamp>.rst
            ├── quality_check
            │   ├── sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>_corr.tsv
            │   └── sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>_maskdiff.tsv
            ├── sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>_affine.mat
            ├── sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>_defacemask.nii.gz
            ├── sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>_maskdiff.tsv
            └── sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>.nii.gz

Description of contents:

• dataset_description.json Metadata describing the process, including versioning and processing information.

• figures/histogram_mean_correlation_<T1w|T2w|FLAIR>.png Image correlation-to-reference image distribution and applied threshold.

• figures/histogram_overlap_<T1w|T2w|FLAIR>.png Image overlap distribution and applied threshold.

• logs/report_<timestamp>.rst Contains group-level workflow steps and parameters.

• log/commands_<timestamp>.rst Contains group-level executed commands.

• quality_check/mask_overlap_<T1w|T2w|FLAIR>.tsv Table containing the overlap score for each subject/session/run. The table includes a binary qc column indicating the quality control result.

• quality_check/mean_correlations_<T1w|T2w|FLAIR>.tsv Table containing the correlation score for each subject/session/run. The table includes a binary qc column indicating the quality control result.

• subjects/sub-<id>/ses-<id>/figures/sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>_defacemosaic.png A visual mosaic showing defacing masks on some slices for quick quality check.

• subjects/sub-<id>/ses-<id>/logs/report_<timestamp>.rst Contains subject-level workflow steps and parameters.

• subjects/sub-<id>/ses-<id>/log/commands_<timestamp>.rst Contains subject-level executed commands.

• subjects/sub-<id>/ses-<id>/quality_check/sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>_corr.tsv Table containing the correlation score.

• subjects/sub-<id>/ses-<id>/quality_check/sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>_maskdiff.tsv A table containing voxel counts and physical volumes (in mm³) for the brain/defacing masks and their intersection.

• subjects/sub-<id>/ses-<id>/sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>_affine.txt Affine transformation parameters (12 DOF) used to align the T1w image to the MNI template or rigid transformation parameters (6 DOF) used to coregister the T2w or FLAIR image to the T1w image.

• subjects/sub-<id>/ses-<id>/sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>_defacemask.nii.gz The binary mask identifying voxels removed during defacing (face and ears).

• subjects/sub-<id>/ses-<id>/sub-01_ses-01_run-01_mod-<T1w|T2w|FLAIR>.nii.gz The final defaced T1w, T2w or FLAIR image with facial/ear structures removed.

5.5. Featured examples

Defacing

Explore how to perform this analysis.

../auto_examples/workflows/plot_defacing.html

5.6. References

[1]

Fidel Alfaro-Almagro, Mark Jenkinson, Neal K Bangerter, Jesper L R Andersson, Ludovica Griffanti, Gwenaëlle Douaud, Stamatios N Sotiropoulos, Saad Jbabdi, Moises Hernandez-Fernandez, Emmanuel Vallee, Diego Vidaurre, Matthew Webster, Paul McCarthy, Christopher Rorden, Alessandro Daducci, Daniel C Alexander, Hui Zhang, Iulius Dragonu, Paul M Matthews, Karla L Miller, and Stephen M Smith. Image processing and quality control for the first 10,000 brain imaging datasets from UK biobank. Neuroimage, 166:400–424, February 2018.

[2]

Amanda Bischoff-Grethe, I Burak Ozyurt, Evelina Busa, Brian T Quinn, Christine Fennema-Notestine, Camellia P Clark, Shaunna Morris, Mark W Bondi, Terry L Jernigan, Anders M Dale, Gregory G Brown, and Bruce Fischl. A technique for the deidentification of structural brain MR images. Hum. Brain Mapp., 28(9):892–903, September 2007.

[3]

O. F. Gulban, Nielson D., and Poldrack R. Poldracklab/pydeface: v2.0.0. 2009. URL: https://zenodo.org/records/3524401.

[4]

Andrew Hoopes, Jocelyn S Mora, Adrian V Dalca, Bruce Fischl, and Malte Hoffmann. SynthStrip: skull-stripping for any brain image. Neuroimage, 260(119474):119474, October 2022.