I would generally suggest binarising by thresholding at 0.5, but there are
other approaches - with their associated subtleties.

For example, consider a case with three possible classes: "GM", "WM" and
"other".  If a voxel is 30% likely to be "GM", 25% likely to be "WM" and 45%
likely to be "other", then how should it be binarised?  Simply thresholding
at .5 would mean that it doesn't belong to any class.  Taking the class with
the greatest probability would put it in the "other" class - while it has a
55% probability of being brain.

I don't think there is any absolutely correct answer, so I would personally be
more inclined to simply use the original belonging-probability images, than
thresholded versions.

Best regards,
-John

> you should keep in mind that the values in the segmented images are
> reflecting the probability of being grey matter. So, if you run a formula
> of i1>0.01 means, you take all voxels with a probability of being grey
> matter of 1%, that's rather low.
>
> So, it depends, what are you going to do. I'm running some analyses tools,
> where I need binary grey-matter images, as well. There I usually use 30% or
> 50% (or even more), i.e. i1>0.3 or i1>0.5 etc. But, this threshold
> crucially depends on the quality of the segmentation, i.e. the quality of
> your original T1 images. In my case, I'm running a sequence (on a 3T
> scanner) with a very good grey-white matter differentiation, which allows
> me to go to even higher thresholds.
>
> So, I would start with 30%, i.e. i1>0.3
>
> Perhaps, John has other recommendations.
>
> Good luck,
>
> Karsten
>
>
> --
> -------------------------------------------------------------
> Karsten Specht, PhD
>
> Department of Biological and Medical Psychology
> & National Competence Centre for functional MRI
> University of Bergen
> Jonas Lies vei 91
> 5009 Bergen
> Norway
> Tel.: +47-555-86279
> Fax: +47-555-89872
> [log in to unmask]
> http://fmri.uib.no/
>
> > I'm interested in the optional binarization step in VBM John mentioned in
> > Human Brain Function 2nd ed. It says,
> >
> > "A further possible step after segmentation would be the binarization of
> > the resulting tissue class images. Many tissue classification methods
> > produces images where each voxel is the a posteriori probability that
> > that voxel should be assigned to a particular tissue type according to
> > the model. These probabilities are values between zero and one.
> > Binarization would involve assigning each voxel to its most likely tissue
> > class."
> >
> > I have been trying to implement this step of binarization, but I'm
> > struggling with setting threshold for this binarization. This is because
> > non gray matter regions have some values other than zero (such as 0.01,
> > 0.03 etc), so simply performing Imcalc i1>0 results in covering other
> > regions than gray matter. I change the threshold manually right now (eg
> > i1>0.04, which seems to be pretty good to me), but I think there has to
> > be some good formula to set the threshold to extract just gray matter.
> > And also I'm wondering if there is a way to compare the segmented image
> > with the binarization image and see whether right regions are binarized
> > or not. Could you please let me indicate how to do that? I would be
> > grateful for any suggestions.