Dear Dr Moore,
>I am running matlab-spm99 for processing of my fMRI data. If anybody can
>inform as to what extent threshold (in terms of voxels) you choose and the
>reasoning behind this choice would be greatly appreciated. You
>obviously get more diffuse activation w/ a lower extent threshold and more
>concentrated, less total activation, w/ a higher extent threshold. I am
>not too sure about the logic that goes into this decision, especially for
>myself. Any help, comments, or suggestions would, once again, be greatly
>appreciated.
If you set the extent threshold to zero (which I would always do when
initially examining contrasts), then you may well get the clusters
which you believe to be biologically relevant, and then a smattering
of little individual dots on the map which you believe to be noise.
When you come to produce a figure for publication, therefore, you
might choose to set an extent threshold of, say, 5 voxels to get rid
of the little dots, purely so that the figure looks prettier.
I am not sure that any formal logic goes into this decision, but if
one was forced to justify it, I guess that answer would go something
like this:
"I have come to this experiment with an 'a priori' assumption that
any real biological effect will be spread over more than 5 contiguous
voxels. Partly this is because I believe that most cortical 'areas'
in the human are at least a few cm across, and under most
circumstances I imagine that all or a substantial part of any 'area'
will be active at once in my particular experiment. Obviously there
are some experiments where this might not be a reasonable assumption,
e.g. if you repetitively stimulated one very small part of the visual
field, then you might get tiny, but highly significant clusters in
retinotopic areas in prestriate cortex.
Also I have used smoothed data, with a kernel of at least three times
my voxel size, so any very highly significant activation, even one
which is really confined to a single voxel, will produce a blurred
cluster of many voxels, further protecting me from eliminating
biologically relevant clusters."
Strictly speaking, one should make any decision about setting a voxel
extent threshold prior to looking at the data. This is to prevent
you from eliminating clusters which you don't like, post hoc, by
fiddling with the extent threshold. This would obviously not be a
valid thing to do, and could be seriously misleading.
If I expect, a priori, activation in tiny areas like the
Edinger-Westphal nucleus or something, then obviously I won't be very
keen on applying the 5 voxel limit. But if I found an unexpected,
highly significant cluster of fewer than 5 voxels, I personally would
assume that it is either some sort of artifact (as can be produced at
the surface of the brain or at the edge of a ventricle by motion) or
that it is just a false +ve (even highly unlikely false +ves have to
happen sometimes!).
If, after the event, a cluster of fewer than 5 voxels makes
biological sense, then perhaps you would wish to report it. If so,
and prior to the experiment you had decided that your published
figures would have a 5 voxel extent threshold, then in theory you
have broken the rules. But perhaps this is being unduly harsh (I
suppose it would depend on how many voxels in the whole brain would,
post hoc, have a 'biologically plausible' reason to be activated. If
the answer is 'loads of voxels', then again you could be seriously
misled).
Sorry, no great insights here,
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]
|