Hi again
On 30 Dec 2006, at 20:42, Christina Hugenschmidt wrote:
> Hello again! I apologize for the barrage of novice questions about
> ICA.
>
> I see from re-reading posts about group analyses that I should
> apparently
> either be:
> 1. concatenating all my data from each group into one 4D file
> ("the current version of melodic does not perform group ICA - for now
> you could try concatenating all data into one 4d file and run
> standard melodic on this file. After estimating the spatial maps
> and time
> courses you can split up the latter into subject-specific chunks and
> feed these timecourses into some meta-analysis.)"
>
> or
>
> 2. using the information in the melodic_IC file, which is the
> unthresholded
> z-map data. ("Once individual Z-maps have been generated it is a
> question
> of *how* you combine these. . . you'll need to model those variance
> components separately. There is no fundamental difference in the
> required
> statistical procedures between ICA and general linear
> modelling"; "unthresholded Z-maps (the maps contained in melodic_IC)")
>
> This leaves me with two questions:
> 1. If I concatenate all the data from one group of subjects into
> one 4D
> file, I do not understand how to then split the time course into
> subject-
> specific chunks. It seems like when you put everyone into one file,
> you are
> treating your group of subjects as an individual. How do you recover
> individual subject information from this in order to account for
> between-
> subjects variance?
>
The final output file melodic_mix is just a text file, e.g. if you
have 10 subjects of 100 EPI volumes each and melodic estimates 20
components then this is a text matrix of size 1000 by 20. Simply load
into matlab (load melodic_mix -ascii) and split this matrix up.
> 2. If I choose to evaluate components from the melodic_IC file, how
> do I
> find individual components in this file? It appears to be just one
> file,
> unlike the separate files for each component in the stats folder. I
> looked
> in the header, and that looked basically the same as the headers for
> individual components in the stats folder.
>
You can load this file into fslview and look at the different
volumes. If you want to extrcat a single component map from
melodic_IC you can use 'avwroi melodic_IC outfile number 1' .
Alternatively, avwsplit melodic_IC splits the 4D file into different
volumes.
> I also have a couple of questions relating to a mistake I made.
> Let me preface my questions with the fact that I am rather abysmal
> at math
> and I am playing with this on my own. It seems to me that if you
> can create
> a spatially normalized z-map of a given component, you should be
> able to
> determine your component of interest for each subject, and then
> perform
> group statistics on these z-maps in a regular GLM analysis.
>
Yep, you can simply concatenate all the volumes from your individual
ICA runs and then run Feat on this:
avwmerge outfile map1 map2 ... mapN
> I did not realize that the z-maps in the stats folder output by
> melodic
> were all thresholded. I picked one component for each person and
> entered
> these into two-sample t-test in SPM. If this is already giving you a
> headache, all apologies.
>
> When I did this, I found that the first component for each person was
> different than the other components. In the header of the first
> component,
> aux_file=none, and the cal_min and cal_max values varied. In all
> the other
> components, aux_file=render1, cal_min=-0.95, and cal_max=1.05. All the
> other variables in the header were the same. Why are these variables
> different in the first component? What is the aux_file?
>
This should not matter - the aux_file is what we use to indicate to
e.g. slicer that this map can be rendered in a particular way.
> Secondly, I have run ICA multiple times on one subject. I got a
> different
> number of components, so I ran the analysis again, this time
> without even
> closing the window, just hitting go again after the first analysis
> finished
> (in case I had mistakenly changed something between analyses). This
> time I
> got the same number of components, but the components were
> different. Why
> is this? If I am inputting the same data with the same settings,
> shouldn't
> the output be the same?
>
I suspect that you did not switch off pre-processing for subsequent
runs? If you run melodic multiple times on the same data the number
of components should not change. However, the order of components
might very well change. Also, there will be some run-to-run
variability - ICA is an iterative process which hopefully converges
to the same solution but ICA results can differ depending on intial
starting conditions, the exact algorithm used etc.
> I cannot thank the list, and Christian in particular, enough for any
> insights you can offer. Happy New Years!
>
Happy new year to you, too
cheers
christian
> Christina
--
Christian F. Beckmann
Oxford University Centre for Functional
Magnetic Resonance Imaging of the Brain,
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
Email: [log in to unmask] - http://www.fmrib.ox.ac.uk/~beckmann/
Phone: +44(0)1865 222551 Fax: +44(0)1865 222717
|