Hi Steve,
I had so much faith in MELODIC...as a broken man I'm now obliged to ask twice.
No, :) it was my first time using the mailing list, I was expecting a copy of the message sent from my mail client. When I sent the post from the website and didn't get one I sorted the archives and saw my message twice, sorry.
Anyway, thanks for the info, it is of great help.
-Prantik
________________________________________
From: Stephen Smith [[log in to unmask]]
Sent: Monday, November 16, 2009 4:24 PM
To: [log in to unmask]
Subject: Re: varying ica results from exec to exec separated by seconds, reproduces across platforms
Hi - seems like you sent the same message twice to the email list - were you testing whether you would get different answers if you submitted the same question twice.....?
Talking of which.....yes this is a well-known issue. ICA is generally initialised with a random initial mixing matrix (unless your ICA software always initialises it to the _same_ 'random' matrix.....which some software does.......very naughty....). So yes, you can get different results each time you run MELODIC, which often tells you more about the data than the algorithm, namely that the different solutions have similar ICA cost functions, i.e. that neither is strongly 'better' than the other. There are things you can do to help tend towards a 'stable' solution, such as run ICA several times and then combine somehow across the various solutions (e.g., meta-ICA, ICASSO), but these solutions may hide the fact of the various solutions you might have seen in the various initial runs.
Cheers, Steve.
On 16 Nov 2009, at 18:32, Kundu, Prantik (NIH/NIMH) [F] wrote:
Hello
I've run MELODIC on a few datasets now to test reproducibility between multiple execs on my machine and multiple execs on another platform, and I find that there is serious variability, from one result to the next obtained within 30 seconds of each other.
Running MELODIC from the GUI, I create a configuration and disable slice timing correction, volume registration, blurring, and spat norm (I have done these externally), and retain the highpass of 100s. I hit Go. Browser pops up and report is generated live as usual.
Then, without closing the MELODIC_GUI, I only change the output directory to something like 'trial2', and hit 'Go' again around 30 seconds after.
I view the output simultaneously.
Prestats are identical, and PCA eigenvalue curves are the same.
Comparing between the two execs, usually the first 10 or so components are comparable but in different orders.
The components that are similar have variances percentages that can easily be 10% different from one another. While they are roughly similar, the timecourses and power spectra have differing subtle features, and spatial localization maps are different too.
A few 'neural' components are within the first 10, most are motion/artifact.
As we go to the lower components we find more 'neural' components, but where some components are clear and pretty in one decomposition, they may be completely absent or heavily intermingled with high freq noise in another.
I tried this on my Mac, then I tried this on a linux box. Same effect. What's going on?
-Prantik
---------------------------------------------------------------------------
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director, Oxford University FMRIB Centre
FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
[log in to unmask]<mailto:[log in to unmask]> http://www.fmrib.ox.ac.uk/~steve
---------------------------------------------------------------------------
|