FMRIPREP: images not aligned to MNI coordinates


After using fmriprep on our ~130 participants’ data, we found an odd issue with the functional scans of very few (4) of them. For some reason these participants’ functional scans were normalized to MNI, but not aligned to the correct MNI coordinates. So the outcome images are of normalized brains in an offset position.

See for example the attached images of the final MNI space preproc images:

(MNI brain mask overlaid in yellow).

The report html files look OK.

Any of you had similar issues?
Any Idea why this happened, and how to solve this issue?


Could you share the HTML report for one of the problematic participant?

If you could share the output files as well (just the preprocessing in MNI) that would be also very useful.

Sure thing!

Here is the preprocessed image and the html file:

Thank you so much, Chris!

Is that happening in a particular session/run or it is participant-wide?

It happened to all functional scans of specific participants.
The anatomical scans were actually registered perfectly well.

Ok, I’m checking out the html report of sub-008 that you shared. The EPI-to-T1 registration seems OK for all runs, is that correct?

I had a look at the data you sent and indeed reslicing did not work correctly. I opened an issue: To help us fix this it would be great if you could provide (here or under the github issue) a minimal set of data we could use to replicate this (for example just the _T1w images and one _bold image). Thanks in advance!

PS @Tom_Salomon I noticed some banding artefacts in sub-008_ses-02_task-probe_run-02_bold run. I recommend looking into this using MRIQC ( now also available on</shameless plug> :wink:

Thank you very much, Chris and Oscar! Super appreciate it! :smile:

I added the original unprocessed data to the previous google drive link. Hope it will be useful

The original data contained two session with about 4G of data per subject, So I added both :

  1. a subset of the data - just the anatomy scan and one functional scan (named it ‘unprocessed_data_sub-008_partial’),
  2. the entire 4G, in case its needed to reproduce the result. I assume fmriprep probably used both sessions to perform the registration to MNI space, so you might need to use it all.

We’ll definitely check out MRIQC next. There’s a lot of data to go over with this experiment…

Awesome - thanks!

MRIQC is great for sifting through lots of data to look for poor quality outliers. Check out this example group report (click on dots to see individual reports).

Hi @Tom_Salomon, we are having trouble with the latest version (1.0.0-rc4) to replicate this problem on your dataset. Could you report which version you were using?

Hi @oesteban. I used v1.0.0-rc2, which was the latest at the time I dealt with this data (several weeks ago). I would try using the latest version as well, though optimally I would prefere if all of the subjects could be processed with v1.0.0-rc2.

Right now we are in the middle of a testing sprint, so I would strongly recommend using the final 1.0.0 version when it arrives, since all these problems are to be addressed for that release.

Hi @Tom_Salomon, I’m still unable to replicate this (

Can you provide more details?. Particularly, what are the options you are setting in your command line?

Thanks very much!

Is there any chance you’re using a pre-computed FreeSurfer reconstruction? It’s conceivable that the transform correction we use to map between FreeSurfer and the “native” T1w space could be affected by our template creation.

Thanks guys.
Interesting. So maybe it was just a temporary glitch and running these subjects again should do the job?

We were actually using pretty much the standard command using docker:
docker run -ti --rm -v /input path:/data:ro -v /output path:/out poldracklab/fmriprep:latest /data /out/out participnts

I think we added to this default command a use of a working dir and no fieldmaps option only.
Without using any prior freesurfer outputs or anything like that.

I am out of office for the next couple of weeks. @Rotem could join in the conversation and let you know if running these subjects again replicated the issue.

1 Like

Hi @Rotem, @Tom_Salomon,

We were discussing this issue again, and it came up that one situation that can result in misaligned outputs is running a dataset with an old working directory. If you’re still having issues, it may be worth trying running again with an empty working directory.

1 Like

Thank you very much for the thorough examination!
We ran some of the participants again with the same fmriprep version but a new working directory and it indeed worked!

Thank you very much for all your help!