Hi,
Well, the entire point of smoothing is to change the SNR characteristics of your data - that’s what you see in action. With sensible levels of smoothing more signals become detectable. Also note that melodic is designed to estimate the dim conservatively (i.e. prefers to estimate more rather than fever components). And indeed, how you smooth (AFNI or FSL) should not make much of a difference
The run-time scales with the number of components, i.e. it’s not the smoothing per se, it’s the higher model order that increases run time.
hth
Christian
On 09 May 2014, at 02:39, SUBSCRIBE FSL Anonymous <[log in to unmask]> wrote:
> Hi there,
>
> I am puzzled by this and am hoping someone here might understand it:
>
> When I do single-session Melodic ICA on unsmoothed 20-min datasets (previously motion corrected, slice time corrected, BET brain extracted, and detrended), I get <<100 components with the automatic dimensionality estimation. However, if I smooth the data, either before Melodic or as part of the prestats options in the GUI, I get 400-600 components. I have done the smoothing with both FSL and Afni, and get the same result. I'm also noticing that Melodic runs much faster on unsmoothed data than on smoothed data.
>
> Can anyone see a logical reason for these observations?
>
> Thanks!
>
> Kajsa
> Princeton University
|