So far, the evidence that I've seen seems to suggest that inter-subject
registration using Dartel tends to be more accurate than the registration
obtained with Unified Segmentation. Certainly, for the comparisons that
we've done here in the FIL, the Dartel results have made more sense than
those using the older spatial normalisation approaches in SPM. I would be
pretty surprised if other people's data did not follow a similar trend.
I would suggest looking at the contrast images obtained using the different
registration approaches. These will show you the general trends in the data.
Then take a look at the ResSS image to see how the variance appears. Maybe
there is an outlier somewhere among your data. For example, if the
segmentation is given poor starting estimates for the positions of the head
in the field of view, then it will mess up.
Best regards,
-John
On Monday 13 October 2008 13:53, Marie-José van Tol wrote:
> Dear list,
>
> Recently, at our lab we switched from optimised VBM to DARTEL to perform
> spatial preprocessing in voxel-based morphometric studies, as it should
> give cleaner results. However, we have noticed that replicability of Dartel
> vs. optimised VBM is not too good in several data sets in which we compared
> samples of psychiatric patients to healthy controls (there was a previous
> post http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0806&L=SPM&P=R15164 for
> Parkinson patients vs. controls). Specifically, we have two fairly large,
> independent samples of n=105 (55 patients vs. 50 healthy controls) and
> n=142 (99 vs. 43); in the first we found robust results (at a corrected
> threshold, in expected brain regions) for optimised VBM, which we were
> unable to replicate when using Dartel (with either 8 or 12 mm FWHM
> smoothing). In the second, highly significant results obtained when using
> Dartel could not be replicated - not even remotely - with VBM. In both
> studies, analytical models for VBM and Dartel were identical. The templates
> created by DARTEL (template_6 (both grey and white matter)) look fine, as
> did the normalized grey and white matter images per subject
> (mwrc1xxxx.nii/mwrc2xxx.nii).
>
> Do others have similar experiences? Any thoughts?
>
> Best,
>
> Odile van den Heuvel, Marie-José van Tol & Dick Veltman.
>
> Department of psychiatry, VUmc, Amsterdam, the Netherlands.
|