Dear Peter
I would advise segmenting normalised images (ie normalising first then
segmenting) since the prior probability maps for segmentation have been
normalised (in stereotactic space).(as Darren has pointed out)
WE have introduced an optimised method to remove non brain voxels prior to
normalisation, since these non brain voxels (in scalp, diploic space,
venous sinuses, petrous apex , clivus etc can introduce error(in press
NeuroImage. Good et al: a VBM study of ageing in 465 normal adult human
brains).
We do a fully automated brain extraction (whwich incorporates an initial
segmentation and then masking with the grey (or white) matter image. We
then normalise the masked grey (or white)matter image to our grey (or
white)matter template (which we usually construct to be scanner specific,
and disease, age/sex specific as well). We then apply the normalised
paparmeters (sn3d.amat files) to the original structural images so we can
perform an optimal segmenation of normalised images(using the same
customised templates as priors). A second brain extraction is then
performed to remove any non brain voxels. We then modulate (modulate with
the Jacobian determinants to correct for volume changes induced during the
spatial normalisation step and finally we smooth modulated and unmodulated
images. Analyses of modulated images test for differences in grey (or
white) matter volume wheras analyses of unmodulated images test for
regional differences in grey matter concentration (see Ashburner &
Friston,2000. VBM: The methods, and Good et al, NeuroImage in press: A
voxel-based morphometry strudy of ageing in 465 normal adult human brains).
With this optimised method, it is important to centre the images to the
anterior commissure and correct any major rotations (using the reorient
images facility) since the initial segmenattion to remove non brain voxels
is on native (un-normalised images)
Another point: if you are comparing disease groups and looking for global
atrophy, then you may wish to model total intracranial volume (TIV) instead
of global grey(or white) matter either as a confounding covariate (ANCOVA)
or proportional scaling.
I tend to model TIV to look for global atrophic changes, and then model
global grey matter to look for regionally specific changes within the grey
matter compartment. I think they are complementary.
I would also suggest looking at white matter and CSF changes, since some
grey matter changes may justy represent a boundary shift from the adjacent
white matter or CSF compartment
I hope this helps
Tina
25/03/01 +0200, you wrote:
>Dear SPM ers,
>we plan to compare a group of patients to healthy volunteers using voxel
>based morphometry.
>In the paper by Mummery et al. first a normalization is done and then
>the individuals are segmented. Would it not be advantageous to:
>a: normalize the whole brain
>b: segment the non-normalized brain
>c: using the paramters of (a) to normalize just the grey matter images.
>I would expect a better segmentation before normalization !?
>Any hints and opinions would be welcome.
>
>Peter
>Sabine
>
>
Dr Catriona Good
Clinical lecturer / Neuroradiologist
Wellcome Dept of Cognitive Neurology, ION
12 Queen Square
London WC1N 3BG
email: [log in to unmask]
phone 0207 8337485
|