Going back to MNI space for RSA

[Algebreaker]

Hi all,

I’m running GLMsingle on fMRIPrep-preprocessed BOLD data and want to use the resulting betas for ROI-based RSA with nilearn.
GLMsingle outputs betas as a 2D NumPy array (trials × voxels).

The betas have shape (n_trials, 272426) while the fMRIPrep mask produces only 111905 voxels. This mismatch causes inverse_transform to fail when trying to map betas back to brain space. Am I missing something - is there a masker that is saved somewhere? Or an alternative workaround to work directly on the betas? Apologies if this is quite basic - I'm still trying to wrap my head around some of the outputs.

Thanks a lot for your help, and for making GLMsingle available!

[kendrickkay]

Could you provide some information / code on how exactly you are calling GLMsingle?

[Algebreaker]

Thanks for the quick reply! Sure, I am using the python implementation of GLMsingle:
I am using the **preproc_bold.nii.gz files from FMRIprep and corresponding events.tsv files as my input, i.e. I load the 4D time series with nibabel and return it data matrix shape (n_time, n_voxels). Then I create a single trial design matrix from the events file.

The output I get per subject and session (aside from the figures) is:
DESIGNINFO.npy
RUNWISEFIR.npy
trial_labels.txt
TYPEA_ONOFF.npy
TYPEB_FITHRF.npy
TYPEC_FITHRF_GLMDENOISE.npy
TYEPD_FITHRF_GLMDENOISE_RR.npy

The call is simply:

results = glmsingle.fit(
    run_design_list,   # list of binary single-trial design matrices (time x trials, one per run)
    run_data_list,     # list of 4D BOLD data arrays reshaped to time x voxels (one per run, whole brain)
    stimdur,           # median trial duration in seconds (from events.tsv)
    tr,                # TR read from BIDS metadata
    outputdir=str(output_dir_path) if save_disk else None,
    figuredir=str(figure_dir_path) if save_disk else None)```

[kendrickkay]

Sorry, I am not sure we have enough information to really understand what is going on. It sounds like you load the full 3D volume (preproc_bold.nii.gz) and give that to GLMsingle. GLMsingle will return results for all voxels in the volume. Elsewhere, it sounds like you are doing something with a mask (111905 voxels), which I am guessing is some sort of "brain mask". So perhaps you have to subselect only the results within the mask in order to interact with your inverse_transform thing?

[Algebreaker]

Apologies for the late reply - I managed to figure out where I went wrong in the end; I wanted to apply the fmriprep mask (*_desc-brain_mask.nii.gz) to the GLMsingle results, but did that incorrectly. This can be closed.

[kendrickkay]

Nice! Always good to squash 🐛