Hi,
wrt 3) If you only include the ICs of interest, you no longer correct
for residual motion artefacts, csf, white matter signal and so forth.
Basically, DR then becomes a seed voxel (or seed roi) analysis with no
correction. Think of the melodic_IC as the design matrix in a GLM
(that's actually what it is if you look at the script): removing
artefactual components would equal removing regressors of no interest.
Your results will thus become less specific.
wrt 2) What I sometimes do is provide the final randomise step with a
gray matter mask, generated from the cohort's T1 images, replacing the
DR generated mask (which is quite liberal). We argue that this
basically constitutes an anatomically motivated ROI analysis. Maybe
there are other thoughts about this by some of the experts?
wrt1) Sorry, can't help you with that. fsl_regfilt is always a nice
idea, but very laborious....
Cheers,
Cornelius
On Thu, Oct 27, 2011 at 1:23 AM, Michael Chen <[log in to unmask]> wrote:
> Hello,
>
> I've been trying some group resting-state analyses using tc_ica in melodic, then dual_regression. I have a few possibly silly questions so far:
>
> 1) In the dual regression dr_stage2_ic???? maps, there is often a large contribution of ventricles. Looking back at the melodic output, there is indeed some voxels from ventricles, even in components that should be relatively clean (auditory) maps. Are there ways to reduce this noise? Using a thresholded map as input to dual_regression, or running fsl_regfilt to denoise individual subjects? We let melodic estimate the number of components, would fixing a larger number reduce this problem? This last solution seems somewhat arbitrary.
>
> 2) In our first test of the whole procedure, many voxels in the randomise output appear to be voxels at the very edge of the brain, or at the edge of the somewhat limited FOV of the scan. We ran BET as part of the melodic gui, but is it necessary to apply a less lenient bet, or to mask at some point (other than the mask dual_regression calculates)? Registration seemed okay, although with the blank spaces from our FOV.
>
> 3) Is there any difference, other than computational load, between running dual_regression with the full melodic_IC and running dual_regression with individual components (of interest, excluding the noise components) split from the melodic_IC file? And, in the later case, it is preferable to re-merge the components of interest for a multiple regression?
>
> Thanks in advance for any help,
> -Michael
>
--
Dr. med. Cornelius J. Werner
Department of Neurology
RWTH Aachen University
Pauwelsstr. 30
52074 Aachen
Germany
|