Smooth for native space ROI based analysis?

I am putting together analysis pipelines for both task-based and resting state analyses. I am using fmriprep for preprocessing, obtaining output for both surfaces and volumes in both native and MNI space. Native surfaces and volumes will be used for ROI analyses - extracting activity/connectivity estimates from ROIs obtained in freesurfer and using those values to test hypotheses.

I am planning on smoothing the MNI space volumes and surface, before running regression/denoising. My question is, would there be any reason to smooth the native space volumes/surfaces? Or, since we are going to use individualized ROIs based on each subject’s anatomy, should we leave the native space data unsmoothed, preserving the full resolution of the data.


This really will be analysis dependent. Smoothness will increase signal-to-noise ratio at the expense of spatial resolution. The idea is that neighboring voxels likely have similar BOLD activations but with different noise structures, so by essentially averaging them together, you should keep the same BOLD activation shape but the noise will average itself out. Of course in doing so, you may lose some information about precise location. This trade-off happens regardless of space (that is, MNI vs native). Whether it is worth it will depend on what your are planning to do with your data.

Smoothing is very common and recommended when doing group analyses or when going on to connectivity measures. Smoothing will also help in creating clusters of activation rather than just small localized regions of activation that may not pass cluster size threshold of significance. The primary times that I see smoothing not applied is when one is trying to get functional ROIs from a subject (for example, where is task A > task B) to be applied to another method like tractography, which would warrant precise anatomical locations (see this paper as an example as one that did not do spatial normalization or smoothing on functional data).

Another consideration is that if you do use smoothed data, the kernel size may vary from application to application, depending on what kind of ROIs you are looking at. So, if I were to summarize an answer here briefly, it would be that it’s never bad to have the smoothed data lying around, and try to anticipate what kernel size(s) you will need. Hope that helps!


1 Like