My current understanding of anatomical defacing is that it’s recommended (particularly for data-sharing), as it removes potentially identifying facial features. I’ve played around with several different defacing tools (pydeface, quickshear, mri_deface, etc), and each varies in how much in how much of the face it strips. So long as no parts of the brain are removed in the process, is it fair to assume that a defaced anatomical, regardless of the defacing software applied, will be pre-processed (e.g. through fMRIPrep) without issue? I believe fMRIPrep was developed using defaced anatomical images; are there known instances were fMRIPrep fails due to improper defacing (too conservative or too liberal)?
I asked myself a similar question a little while ago and went on doing some comparison on our data. I used BIDSonym tool (BIDSonym/bidsonym at master · PeerHerholz/BIDSonym · GitHub) to run the different defacing programs.
Except with deepdefacer, I got consistent results for all the other defacing algorithms when looking at MRIQC metrics of the defaced images (some of these metrics being computed after some preprocessing steps such as segmentation). It is a first step showing that the standard software may deal identically with those image, but some additional tests about segmentation, brain extraction in the context of fmriprep would be interesting also to be sure that there will be no surprise there.
Here is a summary of my tests:
Here is a summary of my tests with MRIQC and pydeface, quick share, mri_deface and deep defacer :
In a prospective way, I also compared the processing time and compared the CAT12 quality results for the defaced image. Here is the kind of results I got for one subject:
|algorithm||Running time||CAT12 QA||Cat12 cortical thickness (mm)||Cat12 ITV (cm3)|
|spm||20s (in matlab)||86,74%||2,71||1709|
-SPM de-face (spm tools) :
-FreeSurfer mri_deface: mri_deface - Free Surfer Wiki
@jsein thanks for this!
While there is no “standard”, I think many of these programs work well enough. Atypical populations or acquisitions may not work as well. CT acquisitions have to be handled differently too. This recent paper compares many of the various software packages:
In AFNI, we provide a version that optionally refaces instead of just defacing. This may be useful for some skullstripping or machine learning algorithms.
10.16.1. Using @afni_refacer_run — AFNI, SUMA and FATCAT: v21.0.21
A key component to doing a proper comparison of various defacing algorithms is to employ thorough visual inspection. Even though it seems like a trivial check, I think many small errors are missed because of not paying attention or using a bad tool. I have made a small tool to help with this that may be of use to others interested in a good comparison:
It is very easy to customize or extend, and perhaps a good addition to integrate it with the BIDSonym tool mentioned above. cc @PeerHerholz