open-discussion
open-discussion > RE: Spatial Smoothing before vs. after Compcor
Jul 31, 2015 06:07 AM | Alfonso Nieto-Castanon - Boston University
RE: Spatial Smoothing before vs. after Compcor
Hi Andrew,
Both approaches (smoothing before regression vs. smoothing after regression) are actually equivalent as long as the White/CSF BOLD signals are actually extracted from the original/unsmoothed volumes (as in the Chai et al. paper, and also as implemented in CONN). Mathematically if X is a [number-of-scans by number-of-voxels] matrix containing the BOLD signal values, then the regression and spatial smoothing operations can be characterized as:
H*X*S
where S is a [number-of-voxels by number-of-voxels] convolution matrix implementing spatial smoothing, and H is a [number-of-scans by number-of-scans] projection matrix implementing the regression of aCompCor effects. The order of operations above (i.e. H*(X*S) or (H*X)*S) does not affect the results, so as long as both approaches are using the same matrix H (i.e. they are regressing the same aCompCor components) then both approaches are exactly equivalent.
Hope this helps
Alfonso
Originally posted by Andrew Song:
Both approaches (smoothing before regression vs. smoothing after regression) are actually equivalent as long as the White/CSF BOLD signals are actually extracted from the original/unsmoothed volumes (as in the Chai et al. paper, and also as implemented in CONN). Mathematically if X is a [number-of-scans by number-of-voxels] matrix containing the BOLD signal values, then the regression and spatial smoothing operations can be characterized as:
H*X*S
where S is a [number-of-voxels by number-of-voxels] convolution matrix implementing spatial smoothing, and H is a [number-of-scans by number-of-scans] projection matrix implementing the regression of aCompCor effects. The order of operations above (i.e. H*(X*S) or (H*X)*S) does not affect the results, so as long as both approaches are using the same matrix H (i.e. they are regressing the same aCompCor components) then both approaches are exactly equivalent.
Hope this helps
Alfonso
Originally posted by Andrew Song:
Hi,
As I was reading through several papers on Compcor, I was a bit confused on when the spatial smoothing via Gaussian kernel comes into play.
In 2012 paper by Chai et al, the spatial smoothing comes before aCompCor, whereas in 2014 paper by Muschelli et al, the step comes at the end, after all the regressions have been run.
Although I suspect that there is no 'correct' answer, I was wondering which approach is preferable for what reasons.
Thanks for the help!
As I was reading through several papers on Compcor, I was a bit confused on when the spatial smoothing via Gaussian kernel comes into play.
In 2012 paper by Chai et al, the spatial smoothing comes before aCompCor, whereas in 2014 paper by Muschelli et al, the step comes at the end, after all the regressions have been run.
Although I suspect that there is no 'correct' answer, I was wondering which approach is preferable for what reasons.
Thanks for the help!
Threaded View
| Title | Author | Date |
|---|---|---|
| Andrew Song | Jul 30, 2015 | |
| Alfonso Nieto-Castanon | Jul 31, 2015 | |
| Andrew Song | Jul 31, 2015 | |
| Alfonso Nieto-Castanon | Aug 3, 2015 | |