Fmriprep output badly misaligned?

Hi, I had fmriprep run on a single session dataset (1st time): a T1 and 6 functional scans (each the identical protocol/task). I’ve looked at this data with other tools before, it seemed pretty normal (meaning, I could plot the same post-registration voxel over six scans and see a similar retinotopy stimulus-evoked time course).

This was the command:

fmriprep-docker /Volumes/vision/Data/BIDS/bidsout /Volumes/vision/Data/new_pype/output participant --participant-label sub-01

looking at the html summary output, I see that the first of 6 EPIs is extremely badly misaligned, like nonsense (turned sidewise, rotated 90deg, etc). On that scan I got the message " bbregister refinement rejected". It was rejected also on another scan, which is also poorly aligned but at least in the right universe.

Of the other 4 (bbregister succeeded), three look perfect but one is rather off.

How can I investigate where this went wrong? I used dcm2bids to convert the data this time, could it go wrong there (formatting problem?)?


edit for more detail

What I had done previously that got nice results was a patched-together notebook from this nipype tutorial:

And the inputs to that were niftis that had been converted using the MATLAB dcm2nii function:

So I thought that since the underlying stuff was mostly the same, the dcm2bids → fmriprep would give similar results.

Some more info:
The affine matrices in the (input) niftis all are the same.

These are relatively narrow-angle scans (back of the head).
So I guess that makes the alignment harder.

Here are means of the raw (pre-fmriprep) images:

Here they are after:

To my eye 3 of those are perfect, two are poor and one (the top left) is very bad. How do I track down the problem?

Partial FOV registration is challenging. Which fmriprep version are you using?

You may have a look at this topic and issue therein:

If you are already using the latest fmriprep version, could you try the option --bold2t1w-init header to see if you get some improvement?

1 Like

Thanks! I will try it.

The initial alignment is not too far off, I only wish I could tell the registration not to stray too far from the initial affine. But I understand part of the principle of fmriprep is not setting special variables…

Using version v21.0.2.

It worked! Maybe it was just luck, but all scans now seem sensibly aligned. Thanks again!

1 Like