Hi Marko,
You probably need to include an additional background class so that the
tissue probabilities sum to one at any point. Also, you'll need to
ensure that the logarithms of the tissue probability maps seem to be OK.
There should be no NaNs, and they should appear reasonably spatially
smooth.
I deeply suspect that it won't work so well if you only use three tissue
probability maps. I was thinking about generating more - rather than
less. Eyeball TPMs would definately help, as there is currently a
tendancy to try to include them in the CSF. This means that the
deformation regularisation needs to be a bit on the heavy side in order
to prevent them from being squashed into the brain, so the TPMs need to
be a bit more blurred to deal with less exact registration. TPMs of
GM/WM partial volume would help to deal with thalamus, putamen etc.
This would need a bit of expert knowledge from someone who can decide
which of those slightly ambiguous brain structures should be considered
as GM. A few other non-brain tissue probability maps would probably
also help to differentiate brain from non-brain.
Best regards,
-John
On Mon, 2009-08-10 at 10:57 +0200, Marko Wilke wrote:
> Dear All,
>
> I am playing around with the "new segment" toolbox in spm8 and try to
> understand how to make it use custom priors.
>
> One, I wonder if I can get by with only supplying 3 tissue classes and
> still get meaningful results? Judging from the documentation and some
> hints from the community, spatial normalization may be better as parts
> of the DARTEL idea are implemented. However, as yet I only get dismal
> segmentation results when supplying three classes (on the command line,
> as it seemed not possible via the batch editor).
>
> And two, I noticed the output is in nifti single image format, and I
> found no option in the defaults to alter that. Is there a global option
> to make spm write out results in the "img + hdr" format instead of in
> "nii" format? I have a couple of scripts I would need to modify if this
> were to be the new standard in some pipelines...
>
> Thanks in advance for any insights,
> best,
> Marko
--
John Ashburner <[log in to unmask]>
|