elyb01
commented
4 years ago Hi Shilpi,
Glad to hear you're using these tools!
1) This means any additional denoising you do to your fMRI data should also be done to your nuisance regressors before they are regressed out. The idea is to avoid accidentally re-introducing unwanted variance you've removed from your fMRI data. The main place this has come up for me is removing noise parameters after using multi-run ICA-FIX with movement parameter regression disabled. ICA-FIX involves performing an ICA decomposition of the fMRI data, identifying noise ICs, and regressing them out of the fMRI timeseries. If you then want to perform movement parameter regression, you need to first regress the noise ICs out of the movement parameters before running the cifti_regress_nuisance script using the ICA-FIX denoised movement parameters as nuisance inputs. This generally applies to any other denoising steps you perform; e.g. by default ICA-FIX also performs 2000s high-pass filtering on fMRI data, so the movement parameters should have the same high-pass filter applied. Note that this is only relevant for nuisance regressors derived before cifti_regress_nuisance is run (e.g. movement parameters derived during realignment); if you derive another regressor from your data after using cifti_regress_nuisance (e.g. mean GM signal), variance from earlier regression steps has already been removed. Recent versions of ICA-FIX (after ~2019) included in the HCP Pipelines automatically generate a denoised version of the movement parameters if MP regression during ICA-FIX is disabled; you can feed this directly into cifti_regress_nuisance. I'm not sure if fmriprep outputs something similar, but I recently added a MP_regress_ICAFIX_noise script to the grayordinate_tools package to generate equivalent denoised MPs for use with older versions of ICA-FIX. This should be easy to modify to use on other nuisance regressors you're interested in including. If you have any questions, I'm happy to weigh in!
2) Do you mean the file size is smaller (e.g. 200MB --> 100MB), or the length of the CIFTI timeseries is shorter (e.g. 400 --> 200 volumes)? If it's only the file size, this may be related to the way the data are encoded. I suggest taking a look at the before and after files using wb_command -file-information to see if there are any obvious differences and wb_command -nifti-information
Hope that helps! Let me know if you have any other issues or suggestions, -Ely
shilpitju
Hi,
I tried using your code to regress out nuissance variables from the cifti dtseries file. My fmri data was preprocessed using fmriprep. Please clarify what it means when you say, "Any filtering/detrending applied to the input CIFTI data should also be applied to the nuisance regressors before running."
Secondly, when I created a .mat file of my normalised regressors and run the code, the obtained denoised file is reduced in size (almost half the original cifti input file). Please guide me where I am going wrong.
Thanks and regards,
Shilpi Modi