Darren is entirely correct. The reason for the poor classification of CSF is
that the segmentation model knows nothing about the tissues outside the
brain. This is also why the initial affine registration used by SPM5
segmentation is not very robust.
Anyway, I have extended the implementation of the segmentation algorithm so
that it uses a few more tissue probability images. This is available in the
SPM8b version. I didn't really spend long enough on generating the new
tissue probability maps, so the implementation isn't as accurate as it could
potentially be. However, for most MR sequences that I've tried, it seems to
give more accurate CSF estimates.
Best regards,
-John
On Thursday 29 January 2009 14:43, Darren Gitelman wrote:
> Dear Vincent:
>
> Unless your subjects have tremendous hydrocephalus, since normal
> intracranial CSF volume is 100-150 cc, I agree that this is very high and
> almost certainly not reliable. I have also noticed that values of CSF are
> typically in the 300-500 cc range with VBM (done either with SPM or the VBM
> toolbox). John Ashburner has previously commented on the list that
> segmentation of CSF (at least for SPM5 and previous versions) is not very
> accurate as it might include other tissues such as skull, etc. Other
> messages have mentioned that total intracranial volume is probably better
> approximated by GM+WM than GM+WM+CSF.
>
> Darren
>
> On Thu, Jan 29, 2009 at 8:20 AM, dfwang <[log in to unmask]> wrote:
> > Dear SPMers,
> >
> > I found that the CSF volume calculated by SPM5 is as large as about 1
> > litre, tested on about 80 subjects. Does it mean that this volume is not
> > reliable?
> >
> > Any comments are welcome!
> >
> > Best wishes,
> > Vincent
|