Smoothing (after fmriprep) seems to be done very differently across studies (I’m mostly thinking of resting state data here, but I guess it applies to task based data too) in terms of which kernel size (if gaussian FWHM 5 vs 6 vs 8 etc) or when (before after FIX) or at all.
I’m wondering how ppl feel about other types of smoothing or downsampling procedures to potentially increase SNR, e.g. by using a Hanning filter or increasing voxel size artificially.
As background: we started thinking about this while working on lesion patient data and we’re trying to find a good way to look at single subject data with ICA to identify known rs-networks (or not) in single subject space (so can’t use fix or aroma).
Of course introducing a new parameter introduces another factor to check (e.g. with cluster stability analyses (thanks for the hint Satra!)), but since I’m rather new to the depth of rs-frmi methodology - I thought I crowdsource the wisdom in the neurostar community. And maybe other people are interested in the answers (or thoughts on it) too?
First, there’s no optimal smoothing value.
Oveall smoothing has three effects:
it boosts the SNR --especially so if the signal has (spatially) the size of the smoothing kernel.
It reduces the mismatcgh between different brains, but this is probably not your point
It blurs the shape, size and precise position of activation foci.
So, the best choice, is a tradeoff btw the pros and cons. Sometimes, it is possible some type of cross-validation to choose between alternatives.
If I understand it correctly your suggestion is to go toward supervoxels. This is an idea I would personally recommend. We’ve explored it in detail and have found it was a good solution. A clustering of voxels by Ward’s method is actually good at that. https://hal.inria.fr/hal-01015172
HI Betrand,
thanks for your response! Yup -agreed there is no optimal value, I was just curious what other ppl think so I’m very thankful for your response and will look into your references too!
As an update on what we’re at now (we includes #jmtyszka):
We used flirt to downsample with prior interpolation and it seems that Hanning (apparently the default for sync) is less practical than rectangular since we don’t try to control for aliasing/ringing. We playing with 3 and 4 isotropic now, which seems to be equivalent to Fourier downsampling by cropping.
To expand on what Dorit wrote, we’re looking at downsampling with some sort of low-pass filtering/smoothing in both 3D and cortical surface spaces. Our ICA needs a little better SNR than we have in our current data, so exactly how we approach this becomes key. Your suggestion of using supervoxels is very cool! I’ve used them in the past for tissue segmentation, but never thought of using them for BOLD fMRI downsampling. Thanks for the links - very nice work!
Thx. We try to share as much as possible in Nilearn but can provide unreleased code it needed.
In particular, we have noticed that super-voxels tend to improve ICA outcomes.
B