A likely naive question…
Is there a difference between AFNI’s 3dDeconvolve and the subject-level run in FSL FEAT?
Is 3dDeconvolve meant for task designs where the HRF is convoluted, whereas FEAT cannot account for this when setting up the EVs?
Thanks in advance.
Below are my general impressions:
** The input data in AFNI are usually scaled by voxelwise average across time, thus the regression coefficients can be approximately interpreted as % signal change and are thus comparable across regions, subjects, groups, and studies. The scaling is done differently in FSL (grand mean scaling?), and the interpretation of regression coefficients remains open?
** Slow drifts are typically modeled with Legendre polynomials in AFNI vs high-pass filtering in FSL?
** Serial correlations in the residuals are handled through an ARMA(1,1) structure at the voxel level in AFNI through 3dREMLfit vs serial correlations estimated & smoothed within tissue type in FSL.
For task-related designs, both 3dDeconvolve and FEAT perform time series regression modeling. The regressors for the model is generated through one’s assumption about HRF. AFNI and FSL handle HRF creation similarly although the specific details could be slightly different. You can assume a shape-fixed HRF and create the regressors through convolution. There might have some subtle differences in terms of the assumed HRF type. Alternatively, one may adopt a deconvolution approach through basis functions such as tents and cubic splines in AFNI as opposed to FIR in FSL. In AFNI stimulus onset times don’t have to be locked to the TR grids with the deconvolution approach.
FEAT is kind of like AFNI’s afni_proc.py – does a lot of stuff, producing FMRI data analysis outputs.
- Is there a difference between AFNI’s 3dDeconvolve and the subject-level run in FSL FEAT?
- Is 3dDeconvolve meant for task designs where the HRF is convoluted, whereas FEAT cannot account for this when setting up the EVs?
The answer to 1 is that in general the analyses are similar, but in detail there are many differences in the steps. Gang has mentioned a few of them.
The answer to 2 is harder, since the whole question proceeds partly from a misunderstanding of FMRI data and its analysis. I believe 3dDeconvolve is somewhat more flexible in certain respects than FSL’s linear analysis program (FILM according to FEAT/UserGuide - FslWiki). The quick answer to this question is (a) 3dDeconvolve is meant to set up general linear time series analyses, not just convolved (convoluted is not the right word here), and (b) 3dREMLfit should be used for the solution of the linear analysis setup instead of 3dDeconvolve, and (c) afni_proc.py should be used for the whole shebang because it will make one’s life happy and carefree.
If on is really interested in learning more about using AFNI, then the AFNI Academy videos https://afni.nimh.nih.gov/pub/dist/doc/htmldoc/educational/bcamp_2020_03_aa.html are a place to start – in lieu of the not-yet-resumed AFNI Bootcamps.