> a couple of weeks ago, you gave some references on animal studies
> implementing SPM2 for DBM. I have tried some things, but I am not quite
> happy yet. I followed the easy fix in:
> http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind02&L=SPM&P=R96333&I=-3&X=4
>01D290F7DF34680BA&Y=spm%40fil.ion.ucl.ac.uk
>
> which a sort of works with 4mm smoothing and changing the bounding box to
> [-128 127] [-128 127] [-163 162] (as our current template has a matrix of
> 256x256x326). However I changed the header to show a lower resolution as it
> is in reality. I would like to keep as much resolution as possible. Our
> template will have about 60 microns cubed, maybe better, before smoothing.
> What would be a reasonable smoothing kernel size? Where can I actually
> optimize the code to sample more densily?
The smoothing you use will determine the sampling density. I'm not actually
sure what the best amount of smoothing would be though. This is likely to
need some empirical exploration.
A human brain is about 200mm long and 8mm smoothing is used. I reckon that if
you derive your FWHM by scaling your brain length by about 0.04, then this
may be about right.
Note that I hope to eventually lose the spatial normalise button of SPM, and
do everything through the Segment button. No smoothing is applied to the
images for the segmentation, although it relies on the tissue probability
maps being generated from many subjects/specimins, and therefore a bit
smooth. If fewer example brains are used for the tissue probability maps,
then a little smoothing may be required.
> I tried some hints previously discussed
> (http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind04&L=SPM&P=R350007&I=-3&X
>=646FB73C8D982008F1&Y=spm%40fil.ion.ucl.ac.uk), but either there is not
> enough overlap, or MATLAB is running out of memory. In another trial, I got
> a very tiny pixely image with lots of background, and SPM gave a message
> that no nonlinear warping was perfomed, but just linear.
Too little smoothing results in very dense sampling. I was aware of this
problem, but didn't actually fix it because I hoped to remove the spatial
normalisation.
>
> By the way, could I also use SPM5 to do DBM? I could not find a toolbox.
The segmentation in SPM5 writes *_seg_sn.mat files where the deformations have
had a rigid-body transformation factored out of them using the same
Procrustes methods that Fred Bookstein (second time I've mentioned him today)
uses. Multivariate methods could therefore be applied to the 'Transform'
variable in the headers.
P=spm_select(Inf,'.*_seg_sn\.mat$');
X = [];
for i=1:size(P,1),
t = load(deblank(P(i,:)));
X = [X t.Tr(:)];
end;
There is no multivariate toolbox within SPM5 to do this yet though.
Best regards,
-John
|