Ananth Narayanan wrote:
>> I am certainly not going to try to tell you the math behind it but what
>> it does is that it basically tries to minimize the difference by lining
>> the images up so that they best match each other. For this, they don't
>> have to match voxel-by-voxel. Consider the example that somebody moved
>> between the anatomical and the functional series. In order to use the
>> anatomical for spatial normalization (see also below) it will have to be
>> in register with the functionals.
>>
>
> Thanks for the quick reply. I guess the source for my confusion arises from
> using the term "lining the images up" to "best match each other." So the
> Coreg process stores the parameters in the header. Since it is made to match
> each other, and since the two images are of different resolutions, does that
> mean that there are scaling parameters included in the header too?
>
The header encodes for an affine transformation, so yes it includes
scaling. If you want to take a look at the matrix, you can use
something like
h=spm_vol(filename);
h.mat
>
>> Well, you can estimate the parameters, in which case they will be
>> determined and stored in the header of the image that was not the
>> target. For example, if you coregister the anatomical to the EPI, the
>> necessary shifts will be stored in the anatomical's header *without*
>> having to reslice the data, which avoids interpolation and thus is
>> recommended if you want to normalize the data anyway (as normalization
>> will take these shifts into account). If you do realign and
>> coregistration and spatial normalization, you get away with only writing
>> (=interpolating) the data once.
>>
>
> Secondly, what is the general protocol that one follows in the coreg step?
> When doing the Coreg, you've suggested to coregister the anatomical to the
> EPI. Thus far, I think I've been doing it the other way around. (I've been
> using the T1-anatomical as the "Reference Image," the first volume EPI as
> the "Source Image" and volumes 2-onward as "Other Images.") What are the
> advantages or disadvantages to doing it the way you've suggested? (I assume
> in your method you'd select the 1st volume EPI [or is it the mean image from
> the Realign step] as the "Reference Image" and the T1-anatomical as the
> "Source Image" and leave the "Other Images" field blank?)
>
My personal preference is to coregister functional -> anatomical
(ref=T1, source=mean_functional, others= functionals). I like to AC-PC
align my anatomicals before doing anything else with them, and doing
coregistration in this direction let's me keep my anatomicals where I
want them.
>
>> In case you changed your mind, you can then reslice the anatomical
>> without having to estimate again; incidentally, in the case of
>> coregistration, you would only have to reslice one anatomical instead of
>> hundreds of functionals, so the direction matters.
>>
>> And finally, in case you do not want to normalize in the first place,
>> you may want to coregister and reslice as, for statistical analyses, the
>> data has to match voxel-by-voxel.
>>
>
> On a side note, the T1 weighted anatomical images are significantly
> different from EPI. Whereas fluid is bright in EPI, it is dark in T1, among
> other differences. Does Coreg account for that and compensate for that?
>
That's exactly why coreg uses uses mutual information criteria, which
doesn't need to compensate (it's looking for a joint histogram, not some
minimal difference).
Regards,
Neil
|