Hi
yes, excessive smoothing and/or upsampling will normally increase the
estimated number of dimensions. This is because the estimate of the
source number is based on the Eigenspectrum of the data covariance
matrix: melodic checks for the point from which onwards the
Eigenspectrum smoothly decreases in the way that it would if the data
contained Gaussian stochastic noise only. Unfortunately both
smoothing and upsampling (which also smoothes) render even the minor
part of the Eigenspectrum more 'lumpy' and different from the
expected Eigenspectrum of Gaussian noise. The default setting is that
melodic picks a high estimate rather than a low estimate of the model
order as it's generally better to overestimate than to underestimate
the number of sources. In your case it makes sense to estimate the
number of sources from the raw data first and use this on the smooth
and upsampled data or - as you suggested - upsample and smooth after
estimating the ICs.
Wrt the second question: yes, you can try and represent the overall
set of 'interesting' components as a linear combination of few
generic ICs via a second run of melodic. The rows of the estimated
mixing matrix then show how much each one of the generic components
contributed to each one of the original IC maps.
hope this helps
christian
On 14 May 2007, at 21:44, Timothy Laumann wrote:
> Dear all,
>
> I am attempting to do a group level analysis using Melodic and have
> a few
> questions regarding the appropriate (if any) procedures. I have read
> through the listed posts but still have some lingering thoughts.
>
> 1. I have discovered that if I perform normalization of my
> subjects' data
> (they are partial-brain volumes) to MNI space before using melodic,
> I get
> far more, and less interpretable, components (on the order of 500
> components for a 1000 volume dataset) than if I use melodic before
> any kind
> of normalization (in which case, I get 100-200 components on the
> 1000 volume
> dataset). I have also found that smoothing in addition to
> normalization
> before using Melodic only exacerbates this problem. The normalization
> process appears to be functioning properly based on visual
> inspection of the
> normalized volumes. Why should spatial normalization, or smoothing
> for that
> matter, cause such a dramatic boost in the dimensionality of the
> output of
> Melodic? Does this mean that I should employ normalization only on
> the
> output of Melodic if I want to bring this information to a group
> level analysis?
>
> 2. I am employing an external metric for selecting components of
> interest
> within an individual subject's Melodic output. The selected
> components may
> be anyone and any amount of the components in a given individual. I
> then
> concatenate all the selected components (from the spatial maps in
> melodic_IC) from each subject (the number of selected spatial maps
> varies
> from subject to subject) into a 4D dataset and use Melodic once
> again in
> order to identify components which represent spatial consistency of
> the
> selected components across subjects. Is this method an appropriate
> use of
> Melodic and if so, how can I interpret the statistics that Melodic
> outputs?
>
> Thanks so much! I appreciate any help or advice you can give.
>
> Sincerely,
> Tim
____
Christian F. Beckmann
University Research Lecturer
Oxford University Centre for Functional MRI of the Brain (FMRIB)
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK.
[log in to unmask] http://www.fmrib.ox.ac.uk/~beckmann
tel: +44 1865 222551 fax: +44 1865 222717
|