Re: [Icatb-discuss] Fwd: Artifacts in Components?

[Vince C. <vca...@mr...>]

Thanks for taking the time to reply. The FSL pipeline highpass filters at .15 so yes, I am filtering I suppose. You're entirely right about the N and the number of components; I suppose I'll just have to  chalk it up to experiment differences. 

Sure, I’m assuming you mean lowpass (a high pass would remove most of the relevant info).  Note, in general filtering is not recommended prior to ICA analysis as it removes important information both for separating components and also for more subtle comparisons with variables of interest.  Seems like you are getting reasonable components, but something to keep in mind as you tune your pipeline going forward.

I think Srinivas just answered your most recent questions. ;-)

Best,

Vince

 

On Fri, Oct 31, 2014 at 1:44 PM, Vince Calhoun <vca...@mr... <mailto:vca...@mr...> > wrote:

Hi Lauren,
        This looks pretty good.  Regarding the FNC values, are you doing
filtering prior to computing the correlation? Also, note that Allen and much
of our recent work use a much higher number of components than 30 (e.g. 75,
100, or 150) so that could be why your values are not as similar to those
other studies [note those also were high N (e.g. 603 subjects)]. ;-)

Best,

Vince


> -----Original Message-----
> From: chainhomelow33 . [mailto:ls...@gm... <mailto:ls...@gm...> ]
> Sent: Wednesday, October 29, 2014 10:05 PM
> To: ica...@li... <mailto:ica...@li...> 
> Subject: [Icatb-discuss] Fwd: Artifacts in Components?
>
> Hi Everyone:
>
> I'm using GIFT to perform an ICA with the ultimate goal being a dFNC
> analysis on a couple of session of resting state data. I actually ran my
RS
> data through FSLs MELODIC first so I could confirm the components I got
out
> of GIFT for use in the dFNC and so I ended up using the filtered data that
> came out of FSL as input into the GIFT ICA. So my inputted RS data was
> preprocessed via:
> Removing first 4 TRs
> Slice Timing corrected
> Motion corrected
> Smoothed
> And sampled to MNI space
>
> I ran the data through GIFT with a cutoff of 30 components and what I get
> out looks pretty decent. I got out what I thought to be 5 default mode
> network components plus 15 others that looked to be real. I did a bunch of
> other stuff but to get to the point: After running the dFNC I checked the
> stationary FNC between components via a correlation matrix and the overall
> correlation between DMN components is pretty low -  definitely lower than
> what I've seen in most published work - as are the dFNC correlations
> between DMN components. Looking at the components themselves - it looks
> like my components have included in them parts of CSF or other small
> activations elsewhere that seem to me to be artifactual (read. not
> realistically belonging to the network in question). I've attached a
> picture hopefully giving a good example of what I'm talking about. This,
to
> me, is a pretty clear DMN component but it looks like at the midline
> there's some residual activation as well as in the cerebellum. And these
> are on components that are thresholded pretty highly.
>
> I'm not entirely sure if this kind of activation would mess up the
> correlations a ton or if it's a result of smoothing or what but I want to
> make sure I'm running connectivity analyses on clean components and mine
> are definitely not as clean as your guys' from any of the Allen et al.
> papers or anything else I've read from you. In your experience is this
just
> the nature of the beast or is there a way to ensure cleaner component
> extraction? I obviously have not masked out white matter or CSF but I was
> under the impression that this is not necessary.
>
>
> Any insight would be very helpful. Thanks for your time.
>
> Lauren