> I'm trying to do a simple volumetric analysis of WM, GM and CSF using
> segmentation in SPM2 and have some questions after experimenting a bit with
> it. We find that we get very different results with N vox depending on
> whether the brain is extracted first or not or if we input normalized
> images or not.
If the data is spatially normalised, then you are changing the volume of the
brain, so I would expect different results. If the tissue class images are
scaled by the Jacobian determinants to allow for expansion/contraction in the
warping, then the sum of the values should remain about the same.
Also, in SPM2, try taking a look at what is classified as CSF, but windowing
in around zero. You will find that there are many voxels with a small
probability of being CSF. Summing up these values results in an overestimate
of the amount of CSF. Therefore, for CSF volumes, the brain extraction will
make a difference. If you find differences in the segmentation results
between skull-stripped and not skull-stripped data, then it may be an idea to
try looking at difference images in order to have a clearer idea of what is
happening.
>
> 1. Is it recommended to extract the brain first?
This can definately help - providing the brain extraction doesn't chop of any
bits of brain. Also, don't lose too much of the CSF. If this happens, then
the model may begin to use its extra degrees of freedom (that arise because
it no longer needs to model CSF) to do strange things.
>
> 2. What is N vox giving us? If it is the proportion of intracranial volume
> accounted for by WM/GM/CSF, why don't the numbers add to 1 consistently?
I have no idea here. Where does this "N vox" figure come from? Is it just
the sum of the values in the image? Maybe it is the sum of the number of
values greater than some threshold?
>
> 3. For the most accurate segmentation and estimates of proportions of
> GM/WM/CSF, resp. of ICCV, should the non-normalized MPRAGE or normalized
> MPRAGE be used? Clearly resolution of the resliced, normalized images
> matters as well because of partial voluming and interpolation errors
The answer depends on how accurately the image can be segmented. In SPM2, if a
native space image is segmented, then an affine transform is used to overlay
the tissue probability maps. If your individuals have strange shaped brains,
then this may not be as accurate as overlaying the tissue probability maps on
to the native space images. However, if you segment spatially normalised
images, then there are partial volume effects (and you also need to account
for the Jacobians in order to compute the volumes).
If you use SPM5, then this isn't an issue, because the probability maps are
deformed so that they overlay on to the original image.
>
> 4. Is there a substantial difference between SPM2 and SPM5 with respect to
> segmentation?
I certainly hope so.
Best regards,
-John
|