Gibbs unringing uncertainty

[tsantander]

Summary of what happened:

Hey gang,

Mostly here for a sanity check (which I find is necessary more often than not these days).

I’ve been processing data for a few different studies using a couple different DSI sequences, two with a multiband Q4 sampling scheme and one multiband cs-DSI—both of which I know @mattcieslak is familiar with (sorry to bother you, Matt). We’ve been noticing that the Gibbs-unringing step in every subject produces some curious outputs: namely, the estimated Gibbs artifacts only appear on the front half of the brain—and I really do mean that literally, insofar as there is a sharp cutoff right at the midpoint of the brain along the phase-encoding axis (see images below).

They are both partial Fourier acquisitions (7/8, so using RPG) with P >> A phase-encoding, so I can rationalize to myself why this could be the case from an acquisition perspective, but given how sharp the transition is, I just want to make sure people smarter than me see this as expected behavior. Will emphasize again this happens in every subject, so as far as I can tell it’s not a random issue arising from head motion or other data quality issues.

Thanks as always for the input!

Command used (and if a helper script was used, a link to the helper script or the command generated):

Been running this on different studies (some of which use different fieldmaps) but the command is more or less the same and doesn’t change what I mentioned above:

    /opt/singularity/bin/singularity run --cleanenv \
	    --bind /zwork/tyler/ohs/data:/input \
	    --bind /zwork/tyler/ohs/qsiprepDerivs:/output \
	    --bind /zwork/tyler/ohs/qsiprepWork:/work \
	    --bind /sw/freesurfer/license.txt:/license.txt \
	    /zwork/tyler/qsiprep/qsiprep-1.0.1.sif /input /output participant \
	    --participant-label ${thisSubject} \
	    --work-dir /work \
	    --omp-nthreads 8 \
	    --nprocs 12 \
	    --mem 30000 \
	    --anat-modality T1w \
	    --denoise-method dwidenoise \
	    --unringing-method rpg \
	    --b0-to-t1w-transform Rigid \
	    --output-resolution 1.8 \
	    --hmc-model 3dSHORE \
	    --hmc-transform Rigid \
	    --b1-biascorrect-stage none \
	    --fs-license-file /license.txt

Version:

1.0.1

Environment (Docker, Singularity / Apptainer, custom installation):

Singularity

Data formatted according to a validatable standard? Please provide the output of the validator:

Yes but sorry I had already cleaned out all the SLURM logs—still can confirm qsiprep was correctly identifying all the relevant input data, PE direction from the jsons (the html output invariably says A>>P but from what I understand this may be a bug in the reports), etc.

Screenshots / relevant information:

An exemplar below (but again, happens for everyone):

---

[tsantander]

Sorry for the bump but it seems the plot has thickened since my post the other day. There’s gunna be some stream of consciousness nonsense in here so apologies in advance for that as well.

In any case, I was able to confirm the same pattern of results (after unringing) with the same DSI sequence run across two different sites (both 3T Prisma systems) and on the same Prisma collected years apart (albeit technically different gradient coils).

The call to RPG within qsiprep seems to be perfectly fine (correctly gets the PF from the metadata and appropriately indicates ‘vertical’ phase-encoding), e.g.:

Gibbs /work/qsiprep_1_0_wf/sub_1_ses_1_wf/dwi_preproc_ses_1_wf/pre_hmc_wf/merge_and_denoise_wf/dwi_denoise_ses_1_dwi_wf/denoiser/sub-1_ses-1_dwi_LPS_denoised.nii.gz sub-1_ses-1_dwi_LPS_denoised_unrung.nii 0.8750 1

And it’s also not a plotting bug, as I can reproduce the visual report by manually subtracting: sub-x_dwi_LPS_denoised_unrung.nii - sub-x_dwi_LPS_denoised.nii.gz

So I thought maybe it might have something to do with running dwidenoise first—went and tested out RPG on a couple raw scans (no reorenting to LPS or MP-PCA, just run locally on my machine via the TORTOISE Docker container), and much to my surprise…

It switched! Now the cutoff is a bit more rostral along the PE axis and it’s just the posterior portion of the brain that seems to have estimated/removed Gibbs artifacts. Here’s an example from a cs-DSI collected in May of this year (absolute value overlaid after subtracting the first raw and unrung b=0 volume):

And one using the Q4 sampling scheme (same site but prior to a gradient coil replacement):

Would be thrilled if anyone has any explanation for this. In the meantime, thinking perhaps it’s best to avoid using RPG / skip Gibbs unringing altogether for these sequences?

[mattcieslak]

Hi @tsantander this is truly bizarre. I think we’ll have to check with the author of rpg, but he is furloughed until the shutdown is over.

In the meantime does this still happen if you use mrdegibbs?

[tsantander]

Hey Matt, thanks for getting back to me! And no worries at all, that’s totally understandable. I hadn’t tried mrdegibbs (figured RPG was the right way to go for PF acquisitions), but currently testing out a few scans from different sites/studies.

Waiting on the full pipeline to run so I can compare the final outputs, but for now, definitely looks like it’s an RPG issue.

Here’s an example cs-DSI collected at UCSB:

A Q4 also at UCSB:

And the same sequence but collected at another site:

Much more like what I’d expect! I suppose the (extremely loaded) question at this point is just whether mrdegibbs is appropriate to use for PF DSI data. I know it was designed for full-Fourier acquisitions, but from what I’m reading it’s probably ok? Also have been working with neuroimaging data long enough to know that there is rarely a straightforward answer as to what the ‘best’ preprocessing methods are haha, but I assume some Gibbs unringing is probably preferable to none, and mrdegibbs at least seems to provide something more sensible.

I appreciate your expertise here, so if you’ve got any more thoughts, I’m all ears. Thanks again!!!

[Steven]

Hi @tsantander,

I’ll chime in and say that the mrdegibbs is not unusable for partial Fourier; there has been some discussion on the MRtrix3 forum about this. Documentation currently reads:

Note that this method is designed to work on images acquired with full k-space coverage. Running this method on partial Fourier (‘half-scan’) may not fully remove all ringing artifacts, and you may observe residuals of the original artifact in the partial-fourier direction. Nonetheless, application of the method is still safe and worthwhile. Users are however encouraged to acquired full-Fourier data where possible.

Best,
Steven

[tsantander]

Hey @Steven appreciate the input! Yeah, the sense I had gotten after a bit of googling was that it’s probably still helpful if not optimal—but I’m traditionally more of an fMRI guy so def nice to get some extra assurance from people w/ more experience in this domain