> I’m hoping someone might be able to help us get to the bottom of an issue
> we’ve run into. Recently we have been trying to reprocess our data using
> SPM8. Unfortunately, we have run into a situation where some subjects which
> segmented fine in SPM5 are now failing in SPM8.
>
> The way we do segmentation is we first register the T1 (MPRAGE) to the T1
> template to give the segmentation a better starting estimate. We then
> segment from there. However, in SPM8 the registration has consistently come
> out differently. It seems that in some cases this new registration is
> giving the segmentation a starting estimate that it cannot handle, where in
> SPM5 the starting estimate was fine.
Are you actually reslicing the MPRAGE scans? Maybe with B-spline interpolation?
I've seen a dataset that has been problematic because of this, where
the scan had large regions of zeros in it, where bits of the FOV
contained no information. Usually, the segmentation ignores zeros,
and assumes that they are masked out parts of the FOV. However, if
the data has been resliced with high-degree interpolation, these zeros
will no longer be exactly zero. The segmentation therefore tried to
fit the air class to the approximate zeros, and another tissue class
to the normal background noise.
If this is the problem, then maybe you could just register the images
without actually doing any reslicing. Maybe try windowing the
intensities in the image to see if there's anything strange lurking in
the background.
>
>
>
> Does anyone have any ideas on what would be causing these differences? Is
> the coregistration and segmentation algorithms very different between SPM5
> and SPM8? If not, is there some setting we may need to change? If they are
> different, are there any suggestions you could give so we could successfully
> segment these cases in SPM8?
The basic algorithms are the same. The implementational details are
slightly different though. The older segmentation tried to ignore the
background (because it had no model for skull, scalp etc), whereas the
newer one tries to model everything in the FOV.
Best regards,
-John
|