Hi Guiseppe,
These are two different types of de-meaning:
(i) "melodic does not remove the mean time course - though you can request this to happen using the --remove_meantc command line option
the mean time course in this case is the time course you get from averaging all the different time courses, then removing it from every single time course. So this an average across space for the tx1 dimensional quantity. This process is typically referred to as 'global mean time course removal' and melodic does not do this, unless you request this explicitly on the command line. In my mind it makes outputs more interpretable (though in the Remes et al paper it is shown that there is an issue in the sense that not removing this leads to worse convergence)
(ii) "time courses are demeaned and then subjected to variance normalisation, i.e. scaled by a robust estimate of the voxel-wise residual noise std deviation"
this describes the process of removing from every single time course its mean value, i.e. centering the time courses around 0. All the mean values that you remove in this process jointly form a map of mean intensity values (the mean EPI) so rather than removing the mean time course (the data mean across space) as in (i) this process removes the mean image (the data mean across time). melodic always does this
So in the first case I'm talking about the _single mean time course_, whereas in the second I'm talking about the _different time course means_ ... easy to get these confused but I hope this makes things clearer
best
Christian
On 15 Jul 2010, at 16:10, giuseppe pagnoni wrote:
> Dear Christian (Beckmann),
>
> a very quick question: does the current version of melodic (3.10) perform voxel-wise temporal demeaning? I was reading a recent paper you co-authored (Remes et al, Neuroimage, in press), that said that melodic did not demean the time series up to version 3.09. I browsed the FSL list and found two somewhat conflicting statements about this (posted on the same day, 18 June 2010):
>
> 1) Re: ICA global signal question
> "melodic does not remove the mean time course - though you can request this to happen using the --remove_meantc command line option"
>
> 2) Re: Melodic ICA question
> "time courses are demeaned and then subjected to variance normalisation, i.e. scaled by a robust estimate of the voxel-wise residual noise std deviation"
>
>
> thanks for any clarification (and congratulations for the paper!)
>
> giuseppe
|