Hey Darren,
sounds like u need a new system! I found a new 1Gb 333MHz DDR system with
an athlon 1700+ (all on an EPOX mbo) basically doubled the processing speed
over a Pentium III 733MHz system with 512Mb RAM. Mind you, if you lack the
RAM it will swap out, so you still need a really fast disk and a dedicated
partition (4Gb max for 32 bit systems). I'd bet your system is swapping out
like crazy. If you can, check the memory page "faults" and i/o.
Cheers, Darren
----- Original Message -----
From: "Darren Schreiber" <[log in to unmask]>
To: <[log in to unmask]>
Sent: Friday, May 02, 2003 3:00 PM
Subject: [FSL] Higher Level Analysis Thresholding Wisdom Needed
> I have been running the higher level analysis of my simplest experiment
> now since Sunday morning. It looks like it will still require a few
> more days of processing to be finished (I am using an 867MHz G4 with
> 512MB Ram.) As I have looked at the initial results for my simple
> contrasts (W, B, N,P), I am getting clusters that contain 20,000
> voxels. For instance:
>
> Cluster List
>
> Cluster Index Voxels P -log10(P) Max Z x (mm) y
(mm) z (mm)
> 2 20330 0 48.6
.06 -30 -92 4
> 1 675 0.000699 3.16 5.23
26 20 26
>
>
> I ran this analysis using the default settings. Clearly, these results
> are not terribly useful for analysis. I ran this using the default
> settings -- cluster thresholding, 2.3, .001.
>
> I am having exactly the opposite problems for the more complex
> contrasts (W>B, B>W, N>P, P>N) and am getting no activations.
>
> Since this first group analysis is just an exploratory run to help me
> learn to run FSL, it isn't a big deal that I'm getting these difficult
> to manage results. However, I am going to be doing my next analysis
> under more time pressure and it is a substantially more complex
> experiment.
>
> How should I go about rethresholding my runs? Especially, since it
> appears I am having two opposing problems. With the simple contrasts,
> (W, B, N, P) my understanding is that I should be raising the cluster
> thresholding, right? But, which of the two numbers should I be raising
> to get interpretable contrasts? Or, should I just switch to
> voxel-based clustering? Or, something else?
>
> Also, what should I do about the more complex contrasts (W>B, B>W, N>P,
> P>N) in order to see some meaningful activations?
>
> And, is there a way to simultaneously achieve these seemingly opposing
> goals? Or, do I rerun the analysis with higher thresholds to get
> meaning out of the simple contrasts and lower thresholds with the more
> complex contrasts?
>
> I am not anxious to meander through the parameter space to find
> reasonable settings for the threshold values since there is such a
> long time requirement for my data to run. So the wisdom of experience
> would be especially critical. Also, since I have more pressing time
> problems for my next experiment, any advice on thinking about it?
>
> Should I use the results that I am generating right now as masks for
> the next run? Would that make sense as a solution?
>
> Darren
>
|