Non-linear registration with QSIPrep?

Software Support
, ,
1

Summary of what happened:

Hello the Neurostars Community,

My question is pretty general and not related to a specific bug:
Is it possible to perform a non-linear co-registration of DWI with T1w images using QSIprep?

I thought at first that the --force-syn flag would perform a non-linear registration (using it even if I have a fieldmap), but the output HTML let me think that only a rigid or affine registration was done.

Can anyone help me on that subject ? I would really need non-linear registration for my DWI as they come from subjects with high cortex deformation thus linear registration is not sufficient.

Thanks in advance for any help.

Command used:

apptainer run --cleanenv --containall \
                --bind {bids_dir}:/mnt/rawdata:ro \
                --bind {output_dir}:/mnt/work:rw \
                --bind {fs_license_path}:/mnt/license.txt:ro \
                {qsiprep_sif_path} \
                /mnt/rawdata /mnt/work/output/ \
                participant \
                --omp-nthreads 16 --mem 32000 \
                --output-resolution 2 \
                --skip-bids-validation \
                --force-syn \
                --unringing-method mrdegibbs \
                --work-dir /mnt/work/workdir/ \
                --fs-license-file /mnt/license.txt \
                -vv \
                --participant-label {subject_id} \
                --session-id ses-{session} \
                --stop-on-first-crash

Version:

QSIPrep 1.1.1

Environment

HPC run with apptainer

2

Hi @Tolema and welcome to neurostars!

That is not possible in QSIPrep. You can only choose between --b0-to-t1w-transform Rigid or --b0-to-t1w-transform Affine .

Best,
Steven

3

Hi Steven, thank you for clarifying this quickly.
Do you think I could register non-linearly qsiprep outputs with other software ? or do you think it could be detrimental to already pre-processed DWI ?
If yes, do you have any advice to do so properly please?
Best,
Tolema

4

Hi @Tolema,

I am not sure about other software. Doing non linear registration after qsiprep adds additional data interpolation and creates need to adjust the gradient table, both of which are not ideal. Could you instead warp post processed outputs? E.g. your microstructural maps, or even streamlines can be nonlinearly warped.

Best,
Steven

5

Hi,
Ok thanks ! So, if I understand well your proposition, I could apply non-linear unwarping inside QSIrecon pipeline ? I am going for a connectome reconstruction, but I guess a non-linear co-registration should be done before this last step of connectivity matrix reconstruction, right ?
Best,
Tolema

6

Hi @Tolema,

I meant outside of Qsirecon. So Qsirecon would generate derivatives in unwarped subject space. You’d nonlinearly warp them.

Best,
Steven

1 Like
7

Hi @Tolema , maybe you could try to use SyN-b0 DISCO for better distrotion correction of your dwi image and then a linear registration within qsiprep would do the job?

8

Hi @jsein,

Thanks for your input. I’m already using SyN-b0 DISCO and linear coregistration, but some of my subjects’ DWI scans show significantly distorted ventricles due to cortical atrophy. In these cases, linear coregistration doesn’t seem sufficient.
However, I’ve noticed that nonlinear coregistration for DWI isn’t widely adopted, likely because of the contrast differences between T1w and DWI images. My reference is the MRtrix documentation.
Have you encountered similar issues or explored alternative approaches?

Best,
Tolema