Hi
The dim estimate will use the peak of the Laplace approximation to the
model order for each run - note, however, that that order estimate is
quite flat in a range around the peak, i.e. it in the example below 32
is the peak but anything from 23-43 looks similarly reasonable. Wrt
the ordering: there is a post hoc ordering of components relative to
the amount of uniquely explained variance - again if multiple
components have roughly the same 'strength' then run-to-run
variability might cause IC1 to end up being IC3 or s if you run next
time. Overall I would expect the components to be roughly in the same
order (i.e. the top 10 components appear early on every time - is that
not the case?) If you want to reduce run-to-run variability one way to
achieve this is to select a smaller convergence criterion (you need to
run from the command line), i.e. use --eps=0.00000000001 if you want
to reduce run-to run variability
cheers
christian
_______________________________________________
Christian F. Beckmann, DPhil
Senior Lecturer, Clinical Neuroscience Department
Division of Neuroscience and Mental Health
Imperial College London
Hammersmith Hospital - London W12 0NN
Tel.: +44 (0) 208 383 3722 --- Fax: +44 (0) 208 383 2029
Email: [log in to unmask]
http://www.imperial.ac.uk/medicine/people/c.beckmann/
Senior Research Fellow, FMRIB Centre
University of Oxford
JR Hospital - Oxford OX3 9DU
Tel.: +44 (0) 1865 222551 --- Fax: +44 (0) 1865 222717
Email: [log in to unmask]
http://www.fmrib.ox.ac.uk/~beckmann
On 8 Jul 2008, at 15:32, Eva Kenny wrote:
> Dear FSL Experts,
>
> I wondered if you might be able to help me with a query I have
> regarding
> the independent components from MELODIC output. I have run the multi-
> subject temporal concatenation approach on a group of subjects and I
> get an
> output with 42 components, however if I change the order I upload
> the 4D
> image files in I only get 36 components. In addition if I split these
> subjects into 2 groups I get 45 and 48 components respectively.
> Also, none
> of the components match with each other, i.e. all IC 1's are not the
> same
> and so on.
>
> Would you recommend preselecting the number of components? If so how
> many
> would you select? Or using the eigenvalue graph or the percentage
> explained
> variance values to select a cut off?
>
> Many thanks in advance,
> Eva
_______________________________________________
Christian F. Beckmann, DPhil
Senior Lecturer, Clinical Neuroscience Department
Division of Neuroscience and Mental Health
Imperial College London
Hammersmith Hospital - London W12 0NN
Tel.: +44 (0) 208 383 3722 --- Fax: +44 (0) 208 383 2029
Email: [log in to unmask]
http://www.imperial.ac.uk/medicine/people/c.beckmann/
Senior Research Fellow, FMRIB Centre
University of Oxford
JR Hospital - Oxford OX3 9DU
Tel.: +44 (0) 1865 222551 --- Fax: +44 (0) 1865 222717
Email: [log in to unmask]
http://www.fmrib.ox.ac.uk/~beckmann
|