Parcellation in native or standard space

Hi all,

I currently try to figure out whether it´s better to perform (atlas-based; Schaefer2018) parcellation in native (T1w) space or standard ( MNI152NLin2009cAsym) space, as provided by fmriprep for connectivity / network analyses.

It seems most of the studies transform their atlas to native space (accounting for individual differences, avoiding interpolation etc.), but some also used fmriprep resampled bold to MNI152 and do the parcellation e.g., in nilearn.

My thoughts so far:

  • (subject space) given the MNI152NLin2009cAsym Schaefer2018 version @Tensorflow it should be possible to use an inverse transformation matrix to transform the atlas to subject space, however I wonder whether it´d a problem that epis are currently not resampled to T1 in fmriprep, but keep their original resolution (e.g. 3x3x3 vs 1x1x1 in T1)

  • (standard space) nilearn functionality is neat when it comes to extracting signal from parcels, however nilearns Schaefer2018 refers to MNI152FSL, which seems not to be equal to MNI152 as provided by fmriprep? However, using Tensorflow Schaefer 2018 would solve this issue.

I would appreciate any thoughts, idea, readings whether native space is superior to standard space because of … , or vice versa!



templateflow and not tensorflow of course :grimacing: