ICA-AROMA and variational Bayesian inference

Hello All!

Our team is interested in using Bayesian inference in SPM to model our data at the first- and second-level. SPM recommends that data not be smoothed before Bayesian estimation while ICA-AROMA (which we have used for motion correction) recommends smoothing. How reliable is ICA-AROMA for unsmoothed data? Any suggestions on best practices here?

Thanks for any help!

I operate under the assumption that as long as the AROMA components are calculated from the smoothed data, you can either use fsl_regfilt to “nonaggressively” denoise the unsmoothed data, or enter the noise components into your first level model with the unsmoothed data. I’m looking up when I believe Martin Mennes said this in another conversation chain.

In short, derive the components from smoothed data, but you can apply the components to unsmoothed data.

Thanks @jdkent! Looking at the ICA-AROMA.py script where the motionICs are applied to the data,

if (denType != 'no'):
print('Step 3) Data denoising')
aromafunc.denoising(fslDir, inFile, outDir, melmix, denType, motionICs)

would it be enough to run AROMA with inFile being the smoothed data, and replace inFile here with a separate argument containing the unsmoothed data?

yep, I think that will do what you wish as well, since that function is essentially wrapping fsl_regfilt.