Hi AM,
Melodic uses full random initiallisation so yes, I'd expect run-to-run variability. If you happen to have two components with quite similar 'strength' then it is possible that there is a change in order but I would not expect this to be a massive difference between runs. You can reduce the run-to run variability by lowering the convergence threshold (--eps on the command line).
What is more of an issue is that even with small changes to the model order (number of components) there is the possibility that a component ends up being split - in which case the order can also change significantly in cases where the two subcomponents reflect a very different total variance than the original (combined) component.
hth
Christian
On 1 Dec 2010, at 11:46, Alle Meije Wink wrote:
> Dear all,
>
> I have a group of 18 resting state fMRI data sets, which I have
> analysed with MELODIC (tc-ICA).
> MELODIC 3.10, fsl 4.1.6
>
> The process is fairly standard: start with the native-space EPI
> images, provide a T1 for every subject for standard space mapping, set
> the TRs, highpass filters etc, and GO.
>
> These settings were all stored in an FSF file.
>
> If I type
> $ Melodic design.fsf & Melodic design.fsf & Melodic design.fsf &
> and then press GO in the 3 windows inside 3s, then everything should
> be exactly the same for these runs, shouldn't it?
>
> The output is not the same though (output, output+ and output++). Of
> the 3 sets of components, two of the 3 runs have comparable IC1s, but
> IC1 of the 3rd run is totally different.
>
> That is odd, because I have heard people attribut meaning to the
> position of an IC in the list. That does not make sense then I guess?
>
> There is nothing I can think of to make these analyses more similar,
> and yet their outputs differ.
> Any suggestions?
>
> Many thanks
> Alle Meije
|