ICA-AROMA and Head Motion Regression

Dear all,

do you think it is reasonable to do first ICA-AROMA and afterwards a GLM with Head Motion Parameters (trans x y z, rot x y z, their derivatives and global signal)?


Hi Stephan,

This is a pretty nuanced question and there are several considerations:

  1. For what application (e.g. functional connectivity, task-contrast maps etc.)? Usually for simple task-contrast maps (e.g. SPM application), you don’t need to be as intensive during denoising. Sounds like you are describing a functional connectivity denoising strategy.

  2. Is this resting or task fMRI? This paper (Parkes et al., 2018) is a good reference for resting state, and this is a good reference (Mascali et al., 2021) for task-based. In both of these, ICA-AROMA (with varying degrees of HMP and physiological signal regression) tend to perform well, with some caveats.

  3. Are your scans relatively long and can withstand the loss of temporal degrees of freedom (tDOF) associated with denoising? If not, maybe reduce your HMP parameter count.

  4. Is your subject cohort prone to high motion (e.g. clinical group, children, task-based scans)?

  5. Do you have a lot of subjects that you can exclude some problematic subjects and still have enough statistical power for your analyses?

Also, I noticed you didn’t mention volume censoring, which is a common regressor to include, but is also dependent on your answers to the above questions. Reading those two papers should provide good guidance. Hope that helps.


Hi Steven,

Thanks a lot for your reply!

  1. Yes, functional connectivity denoising strategy
  2. It is resting-state. The paper you mentioned does not test for HMP+ICA-AROMA.
    What I found is that here (Improved motion correction for functional MRI using an omnibus regression model) but do not know if it is well established.
  3. I have 240 timepoints for each subject
  4. Yes, prone to high motion
  5. No, I cannot exclude subjects.