> Is there any reason to avoid using the "affine only" setting in the
> spatial normalization module? I would have thought that would give an
> affine transform that's more general (flexible) than rigid body.
It is more flexible, but this may not really be what you want. Note also that
by default it regularises the registration, which has a slight biasing
tendancy to zoom up the images (because the MNI brains are bigger than real
brains). This can be changed via the UI though.
>
> Two more questions on within-subject registration in longitudinal VBM:
> * Is there any reason to prefer rigid body over affine?
The best model depends on whether or not you expect the voxel sizes to be
correct etc. If the head moves as a rigid-body, then rigid should be the
better model, as the zooms and shears of the affine should be known. If they
are known, then estimating them will only introduce errors. However, if there
are issues with incorrect voxel sizes in the headers etc, then an affine
model may be preferable. If this is the case, because VBM is actually trying
to compare volumes, your data may need to be caligrated in some way.
> * Is there a reason why, when doing rigid body, "coregistration" (which
> I assume is still more general (inter-modal)) is recommended and
> "realign" is never mentioned?
Realign was coded for working with fMRI time series, and is not optimal for
aligning bigger volumes. However, coregistration only registers using
trilinear interpolation, which itself may not be optimal. Coregistration
does allow a wider variety of objective functions to be used though.
Neither option includes an inhomogeneity model, which would be needed for
achieving really accurate solutions.
I'm not actually sure how good least-squares (the realignment objective
function) is for matching serial anatomical scans. This could be crudely
assessed by plotting a histogram of the residual differences between the
images after alignment. If this is Gaussian, then least-squares is probably
a good model. If not, then maybe mutual information or normalised mutual
information would be a better model (the same probably also applies for fMRI
time series, but I don't plan to get into this).
Best regards,
-John
> From: John Ashburner [mailto:[log in to unmask]]
> Sent: Monday, November 03, 2008 7:17 AM
> To: Fromm, Stephen (NIH/NIMH) [C]; [log in to unmask]
> Subject: Re: [SPM] Longitudinal VBM: non-rigid-body affine coreg?
>
> There is nothing possible via the user-interface, but if you can do a
> bit of
> basic MATLAB coding, then a slight tweek to the estimate function of
> spm_config_coreg.m (around line 280) from:
>
> x = spm_coreg(strvcat(job.ref), strvcat(job.source),job.eoptions);
>
> to:
>
> job.eoptions.params = [0 0 0 0 0 0 1 1 1 0 0 0]
> x = spm_coreg(strvcat(job.ref), strvcat(job.source),job.eoptions);
>
>
> should do the trick. I can't promise that it will work well though.
>
> Best regards,
> -John
>
> On Friday 31 October 2008 15:04, Stephen J. Fromm wrote:
> > I'm working out an analysis path for longitudinal VBM in rhesus
>
> monkeys
>
> > using SPM5.
> >
> > Because the earliest timepoint is at a juvenile stage, and the last is
>
> at
>
> > adult, I'm worried that SPM's coregistration routine (via the
>
> "Coregister"
>
> > button) won't be flexible enough, as it uses rigid body transforms
>
> only.
>
> > Is there a way to do coreg in SPM using a more general affine
> > transformation?
> >
> > TIA,
> >
> > S
|