One reason that I can think of is that the tissue probability maps are
interpolated using B-splines, but without the initial deconvolution.
Proper B-spline interpolation is not really suited to images of this
sort - unless (possibly) the B-spline coefficients are all positive.
Obtaining positive coefficients could be done with an (iterative) EM
algorithm. Another approach is to interpolate the logarithms of the
images and renormalize via a softmax. Note that a tiny value needs to
be added to the probabilities before taking a logarighm.
At the time, I figured that the slight smoothing that resulted because
there was no initial deconvolution probably wouldn't matter that much.
Maybe I should really switch interpolation models, (interpolate the logs
and do a softmax of these interpolated values).
Best regards,
-John
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
On Behalf Of Geoffrey Tan
Sent: Thursday, May 03, 2007 4:58 PM
To: [log in to unmask]
Subject: Re: [SPM] Influence of priors resolution over segmentation
Dear Eric,
I am curious about whether the sampling setting you have used in the
spm5 segmentation accounts for the difference in voxel numbers. There
are 1/8 as many voxels in the 2x2x2mm voxel image as the 1x1x1 voxel
image, so if you sample at the same rate, you have significantly less
information in the segmentation process. Lower voxel resolution
generally looks less crisp and the final image is written based on the
priors I think, so I would worry less about the blurriness if I were
smoothing afterwards.
Best Wishes,
Geoffrey
Eric Pierre wrote:
>Dear SPMers,
>
>We've been trying to undrestand to what extent the resolution of the
prior
>probability maps affects the outcome of the segmentation. We performed
2
>segmentations on the same brain using the VBM5 toolbox (by Christian
>Gaser) and a pediatric prior probability maps (courtesy of Cincinnati
>Children's Hospital Medical Center) at 1x1x1mm resolution, then with
the
>same prior maps at 2x2x2mm (created out of the first one using
>spm_reslice). We got rather significant differences in the segmentation
>results, a sample of those results is attached to this post. What we
don't
>understand is the following :
>Since the prior maps are very smooth, how can changing the resolution
of
>the prior maps influence that much? We couldn't even tell the
difference
>between the 2 prior maps. Any explaination?
>
>Thanks in advance,
>
>Eric Pierre
>
>Case Western Reserve University,
>Case Center for Imaging Research,
>Cleveland, Ohio.
>
>
>
|