Dear Jon,
Jon Brooks wrote:
>
> Hello Richard,
>
> Thanks for the response. Given my terribly worded question,
> you made a pretty good attempt at interpretation! Basically
> you confirmed what I thought might be the case i.e. it's a
> tricky problem to attribute signififcant voxels to one or
> other of the peaks.
>
> The problem I face is that I have fMRI data which (I think)
> have 2 distinct clusters in posterior insula and SII, which
> are separated for some conditions and coalesced for others.
> The coordinates of the peaks in the coalesced clusters
> correspond to those of the well separated clusters. I would
> like to know the cluster size around the individual peaks (i.e.
> to somehow split the contributions for SII and post insula).
>
> I guess that setting a higher threshold across all conditions
> until the coalesced regions separated would be one way
> of solving the problem. Alternatively playing around with
> the FWHM of the smoothing kernel might work (?).
Smoothing has two effects relating to activation peaks and cluster size:
1. Smoothing makes clusters broader and peaks less high.
2. If realignment and coregistration don't work perfectly (and they
never do), true activation positions in different images doesn't overlap
properly. Smoothing then helps to let activation overlap at least
roughly. This again raises peaks.
Both effects counteract: You have one optimal value for FWHM and at both
sides of this value effect Nr. 1 prevailes. Unfortunately this optimal
value is different in every case and, more than that, at every position.
So for best possible parameter estimation you have to fiddle arround
with the FWHM (There is no statistical flaw in that: The statistic is
made for the perfect situation, so lowest p values are nearly guaranteed
to be proper.). Luckily you need to do this only if your data is very
sparse or you have special questions like yours.
There is another point to be made: You can't rely on the exact positions
of cluster borders and peaks. And you can't rely on cluster size.
Reasons are
1. Unperfect realignment and coregistration with locally different
spread
2. Smoothing
3. Heuristic selection of voxel significance treshold
A good parameter to be held in mind is the resel size and the smoothing
kernel size as a minimum of spatial resolution.
As a third: If you have conditions with separated clusters, why you want
to divide the coalesced one in another condition? Since the activation
might be different from a simple sum of the other conditions it would be
a twisty thing to enforce a desired pattern. As pointed out, it wouldn't
have that much value anyway.
If you still want to divide the cluster:
1. You can choose a higher voxel significance treshold
2. You can change to corrected significance for the treshold, what is
nearly the same as choosing a much higher significance.
3. You can choose a smaller FWHM.
4. You can choose a smaller pixel size for the resclicing.
>
> Let me know what you think.
>
> Cheers,
>
> Jon.
>
> _____________________________________________________
> Jonathan Brooks Ph.D. (Research Fellow)
> Magnetic Resonance and Image Analysis Research Centre
> University of Liverpool, Pembroke Place, L69 3BX, UK
> tel: +44 151 794 5629 fax: +44 151 794 5635
>
> On Mon, 12 Feb 2001, Richard Perry wrote:
>
> >Dear Jon,
> >
> >>Not sure if this is a sensible question,
> >
> >I must admit that I am not sure what your question means. So I am
> >certainly not going to be able to supply a sensible answer.
> >
> >>but here goes...
> >>On the results page you may have a large blob, which contains
> >>several bloblets each with their own coordinates.
> >
> >Well, several sub-peaks anyway. They all belong to the same 'blob'
> >though. However, if you were to increase the statistical threshold
> >they would split up into 'bloblets' I suppose.
> >
> >>Is it possible to find out what proportion of the total cluster size
> >>is attributed to each of these sub-clusters?
> >
> >Every voxel in the cluster is significant in its own right, rather
> >than because of its proximity to a peak. So in that sense it can't
> >really be 'attributed to' one of the peaks.
> >
> >However (and it's a big however), your data has been spatially
> >smoothed prior to statistical analysis. So some of the
> >experimentally interesting variance in the voxel may really have been
> >acquired from a neighbouring voxel during the smoothing process.
> >This is mathematics beyond my rather primitive level of
> >understanding, but I wouldn't imagine that you could recover from the
> >smoothed data what this proportion of 'interesting' variance might
> >be. After all, the high spatial frequency information has been lost.
> >I imagine that you would have to go back to the unsmoothed data. If
> >you took a particular voxel in your cluster and went back to the
> >unsmoothed data, ran the statistics, and this voxel came out
> >significant (at the uncorrected level say), then this would
> >presumably indicate that the fact that this particular voxel cannot
> >be 'attributed' to any of the sub-peaks, but is significant in its
> >own right.
> >
> >So I reckon the answer is 'no', but I don't know if I have answered
> >the question that you are really asking. Is it possible, from what I
> >have written, for you to rephrase your question? Perhaps you could
> >indicate the nature of the biological question that you are trying to
> >address.
> >
> >Best wishes,
> >
> >Richard.
> >--
> >from: Dr Richard Perry,
> >Clinical Lecturer, Wellcome Department of Cognitive Neurology,
> >Institute of Neurology, Darwin Building, University College London,
> >Gower Street, London WC1E 6BT.
> >Tel: 0207 679 2187; e mail: [log in to unmask]
> >
Hope that helps
Thomas Kamer
--
Dipl.-Inf. Thomas Kamer
University of Bonn
Department for Psychiatry
Laboratory for Psychiatric Brain Research
Sigmund-Freud-Straße 25
D-53105 Bonn
Germany
Tel: +49-(0)228-287-6366
Fax: +49-(0)228-287-6369
[log in to unmask]
For verification of my signature see https://trust.web.de/root.htm
|