Hi,

I certainly would recommend running explicit motion correction on the data, rather than try to motion denoise. Removing components with fsl_regfilt (rather than add them to your design) is conservative provided you select the components on the basis of a clear criterion for noise (such as correlation with estimated motion) rather than base the decision of which component to remove using knowledge of the (non noise) design. In this case the worst that can happen is that you also end up removing some part of the signal along with the noise. You will not inflate the significance of the signals of interest this way.  As such I don't think adding these regressors as confounds in the design is more appropriate - this approach is actually less conservative. 

Wrt DOF there is no issue: fsl_regfilt removes _structured noise_ in a simple (linear) fashion by subtracting the outer product of  single time series and a single spatial map from the 4D data. The _stochastic noise_ remains intact so you will not see a full loss in DOF. Ignoring this aspect therefore is quite in line with how things are typically done, e.g. when ignoring the effective loss in DOF due to high-pass filtering...

cheers
Christian

On 8 May 2012, at 19:52, Jeanette Mumford wrote:

> Hi,
> 
> That will identify components that look like motion.  I'm unsure if that would do much more than just adding the motion parameters to your regression, did you say you read about it in a paper?  I didn't realize somebody formally tested that, can you point me to the paper if there is one?  
> 
> There are probably other components that are noise but not picked up by motion, but I'm unsure how people go about choosing them other than what is covered in this paper.
> 
> As for removing the ICA time series components from your data, I disagree with only modeling them out of the data in this fashion without taking into account the other regressors in your model (task regressors).  At least it is my understanding that this is how fsl_regfilt works.  It is more appropriate to add them as confounders to your first level regression so the other regressors in your model are adjusted for the variability that is present in your ICA components. Plus your degrees of freedom will correctly reflect what has been modeled from the data.  I realize your DF will vary from run to run, but this shouldn't be an issue in FSL.
> 
> 
> 
> Cheers,
> Jeanette
> 
> On Tue, May 8, 2012 at 11:05 AM, Michael Cordell <[log in to unmask]> wrote:
> Hello FSL Experts,
> 
> I was hoping to get some feed back on my attempt to denoise a set of 4D files using melodic and fsl_regfilt. I am trying to remove any melodic component that has a high correlation with motion. So I followed the following procedure:
> 
> 1.      Ran Single session-ICA melodic on all 4D images
> 2.      Ran a script to run mcflirt on all 4D images and the nextract time courses for each of the 6 motion components (xyz rot/trans) from the mcflirt par file.
> 3.      Used a MATLAB script to run the “regress”  matlab function where the motion parts are run against each melodic component. This is repeated for all images/all melodic components.
> 4.      Collected any melodic component with a p-value under .01 for each image
> 5.      Ran fsl_regfilt removing the components found in 4
> 
> 
> I am hoping for any feedback as to whether I am doing this properly or if there is a better method to achieve the same effect. I have read that melodic denoising is the preferred method for removing motion components rather than including the 6 motion components from step 2 into the first level GLM.
> 
> 
> Thanks,
> Michael
>