> >my impression that SPM2 overregularizes less than SPM99
> >(given the default values of both) stems from just practically
> >comparing the normalization procedure of SPM99 and SPM2
> >(using the EPI-template) in the case of a whole brain
> >functional image with a great deal of signal loss in the
> >frontal orbital cortex. This presents a similar problem,
> >because the normalized EPI-Templates contains tissue in this area.
> >In effect the brain normalized with SPM99 'grew' tissue into
> >the non-existent area (approx an area of 5 cm!)
> >This did not happen with SPM2 for whatever reason.
SPM2 should be more heavily regularised than SPM99. Spatial normalisation
involves a tradeoff between minimising the difference between the images
(mean squared difference), and minimising the amount of distortion. The
model for the amount of distortion in SPM2 is based on the "bending energy"
of the warps, whereas in SPM99 it is based on the "membrane energy". Bending
energy penalises high-frequencies more heavily relative to low frequencies,
than does membrane energy.
> >
> >I assumed that the reason for this is the new cut-off value option
> >in SPM2 (the regularization option is present in both).
This probably doesn't have as much effect as the form and magnitude of the
regularisation.
> >But, of course, certainly more has changed between SPM99 and SPM2
> >regarding normalization than just that.
It's mostly the regularisation and having what I hope should be a more stable
initial affine registration. The other thing is probably that the template
image weighting (brain masking in SPM99) is disabled by default. And a final
thing is that the nonlinear component ignores voxels close to the edge of the
field of view because of the effect of smoothing the data. This should also
reduce unnecessary distortion.
>
> Right. I'm guessing it has to do with the energy/penalty function used in
> the normalization algorithm, but obviously John (Ashburner) is the one
> who'd really know.
There's more in Chapter 3 of http://www.fil.ion.ucl.ac.uk/spm/doc/books/hbf2/
> >>Overnormalization is less likely in SPM2 than in SPM99.
> >>There is a cutoff-value in the defaults (which you could
> >>change yourself to an even more conservative value)
> >>that prevents the normalized brain to extend beyond
> >>the area covered by the scans.
> >>Using the EPI template should therefore work ok.
Estimating spatial normalisation parameters from an image that has been
previously resliced can cause problems. This is because of the pattern of
zeros (or NaNs) that is introduced because data in those regions could not be
pulled from the original image.
> >>Use a deskulled anatomical
> >>(but don't use this image for normalisation!
If you have a deskulled template, then normalisation can work better.
Best regards,
-John
|