Question about confounds from the fmriprep

Hi everyone,

I am using fmriprep to preprocess rs-fmri data. I have some questions regarding the confound regressors given in confounds.csv file and how to use the use them, and was hoping someone can clarify these.

  1. I am planning to use the cosine regressors in the confounds file to high-pass filter the data (i.e., use these along with the other confound regressors I am using to de-noise my data). Are these cosine regressors computed in the same way as shown in this tutorial, using the nilearn.glm.first_level.design_matrix ? When I use nilearn, I get 4 cosine regressors with
from nilearn.glm.first_level.design_matrix import _cosine_drift as discrete_cosine_transform
dc_set = discrete_cosine_transform(high_pass=0.0078, frame_times=frame_times)

where frame_times set as frame_times = np.linspace(0, 320, 400, endpoint=False) since for my data I have a TR=0.8 sec, n_volumes=400, making the task duration 320 seconds. high_pass was set as 0.0078 for one cycle in 128 seconds.

However in the fmriprep confounds.csv, there are 3 cosine regressors, which I am guessing due to having non-steady outlier volumes. I did not opt in for --fd or --dvars-spike-threshold options, yet I seemed to have outliers in the outputted confound file. Is there a way to opt out from having non-steady outliers calculated? If not, would using the nilearn computed cosine regressors be the same?

  1. As detailed in the fmriprep paper, the aCompcor components were calculated using the union of CSF and WM masks in T1 space, after their projection to the native space of functional images. Does this also apply to the global signals extracted within WM and CSF masks, as in they were also projected back to the functional space after being extracted within native T1?

  2. I understand that aCompcor regressors were derived from the high pass filtered functional data, therefore the cosine regressors should be included in the confounds file before de-noising functional data (to have the de-noised data high-pass filtered as well)

If a high-pass filter is already being implemented for the functional data ( e.g., in FSL), can’t we just exclude the cosines but still use the acompcor regressors for the confound regression?

  1. This is unrelated to the above and just to confirm that in the derivatives/func folder sub-001_task-rest_desc-preproc_bold.nii.gz would correspond to the reliagned (i.e., motion corrected) functional data, not to the realigned functional data aligned to T1w space?

For output spaces I used --output-spaces MNI152NLin6Asym:res-1 func \ and opt out from the SyN distortion correction.

Thank you in advance!