> To sort out what to do next, my questions now are how does mutual
> information coregistration work? I know that it is heavily based on
> shape (right?) but I can't quite understand how it works from the
> spm5 manual or its references.

It works based on a joint image intensity histogram.  This is a 2D
histogram, where one axis indicates image intensity of one image, and
the other axis indicates image intensity of the other.  If the images
are in register, then the histogram should contain more structure, such
that intensities of one image can be better predicted from intensities
of the other.  Registration involves moving one of the images around
until this mutual information in the joint histogram is maximized.


> How does the coreg process handle  epi distortion / signal dropout
> between T1 and BOLD-fMRI where there is info in one image and not
> in another

It does not handle it so well.  The model is only rigid-body, so if one
image is a different shape to the other because of distortion, then the
alignment won't be so good.  One part of the brain may be registered OK,
but other parts may not be, because of shape differences.  The effects
of dropout on the registration are slightly difficult to predict.  I
would suggest a bit of empirical exploration here.

> How does the coreg differ from just coaligning two binary masks (is
there
> segmentation in the routine)?

Again, this would need some experimentation to answer.  I think it
depends on how accurately the binary masks can be defined - given the
image artifacts.


> If I wanted to normalize using the strucural, why couldn't I just use
> reorient to match up the first epi image volume with a skull-stripped
> structural acquired with the same exact dimensions a few minutes
> earlier and then use a skull-stripped structural to get normalization
> parameters.

You could do this, but I would hope that coregistration would be more
accurate and easy than manual alignment.  Note that if your subject
hasn't moved, and you use the DICOM conversion routines of SPM, then you
should find that the data are pretty much aligned anyway.

> This way I would avoid the problem of epi artifacts with
coregistration.
> Does this logic sound faulty?

If the EPI are heavily distorted, then there is no way to get the images
into register with a rigid-body model.

> Somehow I have to make sure that the epi distortion won't interfere
> with either aligning the t1 to epi or the normalization process,
correct?

EPI distortion is something that would really need to be corrected in
order to get an accurate alignment.

Best regards,
-John