Thanks all for the help. I will wait the fix for the initial
From: John Ashburner [mailto:[log in to unmask]]
Sent: Friday, July 04, 2008 12:03 PM
To: [log in to unmask]
Subject: Re: [SPM] SPM5 segmentation
Geoffrey is entirely correct. The initial registration (and other parts
of the segmentation) are a bit less robust because the model of the head
only includes information about the distribution of GM, WM and CSF.
There is no information in the model about where to expect bone or soft
tissue outside the skull. I am currently trying to generate further
tissue probability maps with a few more classes so that the head is
rather better modelled. This makes the initial affine registration much
more robust, as well as allowing the CSF to be more accurately
identified. One of the issues with doing this is that a higher
dimensional warping needs to be used, such that unusual faces, thick
necks etc do not adversely influence the deformations within the brain.
If a limited registration model with only about 1000 parameters is used
(ie the cosine transform basis functions), then this would reduce the
accuracy with which the tissue probability maps can be overlaid on the
It's only work in (slow) progress, but it seems to help.
On Wednesday 02 July 2008 09:26, Geoffrey Tan wrote:
> Hi Herve,
> Just to add to this, although I'm sure John will reply to this
> eventually. I think part of the problem is that the prior for
> 'non-brain tissue' does not model non-brain tissue ie skull, blood
> vessels, fat, air... very well and the original affine registration
> can be poor because it registers the brain to the prior probability
> John's working on something that models these classes more explicitly.
> I've personally worked around it by doing an initial affine
> registration of the smoothed image to the MNI T1 atlas, and using that
> as an input to spm_preproc for the affine.
> Marko Wilke wrote:
> > Hi Herve,
> >> Using the unified segmentation in SPM5 (with the more recent
> >> updates installed), we noticed that for a large proportion of
> >> subjects (30%), voxels belonging to the skull were misclassified
> >> into the CSF class (see example below).
> > I believe that the CSF segmentation has never been the focus of much
> > attention ;) leading to the fact that the differentiation between
> > external CSF spaces and skull tissue may not be otpimal.
> >> Has anyone else noticed that and are there any ways to improve this
> >> result?
> > I think what you could to is to focus on the ventricles by masking
> > the CSF segmentation with a "center of brain mask", for example by
> > eroding the default brain mask a couple of times. Alternatively, you
> > could use a more aggressive cleaning/skull stripping routine,
> > perhaps even before segmentation, which has been suggested to
> > improve segmentation accuracy in the first place (a recent
> > NeuroImage paper). Another thing perhaps worth looking into is the
> > prior-less segmentation as implemented in Christian's toolbox, which
> > may (or may not) improve results, again combined with a more
aggressive cleaning approach.
> > Other than that, no ideas.
> > Best,
> > Marko