You might upgrade to the latest version of FSL 5.0.6 and see if you get a
similar answer. There have been some changes to the results of the auto
dimensionality estimation across FSL versions (then again, you might get
something similar).
Peace,
Matt.
On 5/9/14, 11:35 AM, "Kajsa" <[log in to unmask]> wrote:
>I'm using FSL 4.1.9 and Melodic v. 3.10 on a compute cluster
>Kajsa
>
>
>On Fri, 9 May 2014 11:31:44 -0500, Matt Glasser <[log in to unmask]> wrote:
>
>>What version of FSL?
>>
>>Peace,
>>
>>Matt.
>>
>>
>>>Hi Matt,
>>>Thanks for helping me understand this. Resolution is 3x3x4 mm, TR is 1.5
>>>seconds, and 920 time points. No concatenation. When I did ICA on a
>>>different, shorter, dataset (3x3x3mm, TR 2s, 196 timepoints), I
>>>similarly
>>>got a 150-300 components with smoothed data and much less with
>>>unsmoothed
>>>data.
>>>Cheers, Kajsa
>>>
>>>On Fri, 9 May 2014 10:56:18 -0500, Matt Glasser <[log in to unmask]> wrote:
>>>
>>>>What is your resolution, TR, and number of time points per run? Did
>>>>you
>>>>use one long run or concatenate across runs?
>>>>
>>>>Smoothing would only decrease gaussian noise, which ICA is trying not
>>>>to
>>>>put into components to begin with.
>>>>
>>>>Peace,
>>>>
>>>>Matt.
>>>>
>>>>>Thanks Christian!
>>>>>
>>>>>Ok, that makes sense. Somehow I had imagined smoothing would have the
>>>>>opposite effect by decreasing noise sources...
>>>>>
>>>>>On to my next worry: From what I've seen people generally get many
>>>>>fewer
>>>>>components, also with spatially smoothed data. My dataset is pretty
>>>>>standard, although the stimuli are very complex (movies). Does anyone
>>>>>have any ideas about why I get so many more components than other
>>>>>people?
>>>>>I'm worried I've made some mistake. Or doesn't this happen sometimes
>>>>>and
>>>>>it's just that I haven't heard of it in the literature? I'm pretty new
>>>>>to
>>>>>this method.
>>>>>
>>>>>Many thanks!
>>>>>Kajsa
|