The RAVENS registration is more precise than the older spatial normalisation
routines of SPM. In SPM, there are only about 1000 parameters to describe
the warps, whereas RAVENS uses millions. Generally, if the residuals are
very smooth, then it could imply that the model is underfitting the data. The
original concept behind VBM was to treat the spatial normalisation a bit like
high pass filtering in an SPM analysis (ie to remove "macroscopic"
differences) so that "mesoscopic" differences could be detected.
Although the introduction of "modulation" (Jacobian transformation) confused a
lot of people, it meant that the values actually had a more meaningful
interpretation, and that a VBM analysis was not simply an examination of
registration error. "Morphometry" is the study of variation and change in
the form (size and shape) of organisms. With the introduction of the Jacobian
transform, the use of the term became slightly closer to its dictionary
definition, as it involved comparing the regional volumes of grey matter.
This also meant that more accurate spatial normalisation could be used.
RAVENS maps have always implicitly preserved the tissue volumes in the warped
images.
I eventually got around to improving the inter-subject registration model in
SPM. In the latest updates of SPM5, there is a DARTEL toolbox, which can be
used to obtain more precise inter-subject alignment for VBM studies.
Experience of DARTEL in the FIL is generally positive. VBM results produce
higher t stats for the difference that are detected than for the other SPM
spatial normalisation approaches. I would expect the RAVENS results to be
similar.
Best regards,
-John
On Sunday 30 March 2008 09:54, Nikolaos Koutsouleris wrote:
> Dear SPMers,
>
> recently, I tried the HAMMER normalization algorithm of Davatzikos et al.
> on data that had been segmented with SPM5. I observed that the smoothness
> of the obtained RAVENS maps is significantly lower compared the modulated,
> low-dimensionally normalized volumes of SPM5 (see example of the same
> subject in the attachments). Correct me if I am wrong, but this may be due
> to the fact that the HAMMER algorithm uses high-dimensional elastic warping
> which retains a higher degree of anatomical information compared to the SPM
> normalization.
>
> When I did cluster-level statistical inference on the 10mm-smoothed RAVENS
> maps (using Satoru's non-stationarity correction toolbox), I noticed that
> the resulting smoothness estimate was about 5 mm, compared to the
> SPM-normalized data which was about 11 mm. In the RAVENS analysis a primary
> threshold of p<0.05 together with FWE-corrected extent threshold of p<0.05
> produced a minimum number of 1500 voxels, whereas the extent threshold for
> the SPM -normalized data was around 20000 voxels. This is a dramatic
> difference which make clusters significant at a much lower spatial
> threshold. So, I wonder if this difference is really due to the "roughness"
> of the RAVENS maps or if the smoothness estimation in SPM is not valid for
> RAVENS data. Does anybody have similar experiences or possible
> explanations?
>
> Thanks in advance and sorry for this lengthy email!
>
> Cheers,
>
> Nikos Koutsouleris
>
> Imaging Workgroup
> Department of Psychiatry and Psychotherapy,
> Ludwig-Maximilians-University,
> Munich
|