Just to add to Alex Leff's reply (because I had started to write
something and didn't want to bin it)....

| SPM seems to do spatial normalization by warping the BOLD scans directly to
| a template scan in Talairach space (say, an EPI to an EPI template).
| Another method to do this is to map an "anatomical" scan to an "anatomical"
| scan in Talairach space, using say a piecewise linear mapping, then pulling
| the BOLD scan along with the same mapping.
|
| What are the pros and cons of each of these two general approaches? (A ref.
| to the literature is fine.)

I guess the latter case refers to manually picking landmarks on the anatomical
scans, and moving the landmarks to their appropriate locations in Talairach
space with some kind of interpolation between the landmarks. Both approaches
have their pros and cons, but the main advantage of the manual approach is
that human intelligence and knowledge of neuroanatomy can be used, which is
particularly useful when spatially normalising lesioned brains. However,
there are a number of disadvantages of the manual approach, including:
1) The exact locations of the landmarks are subjective.
2) There are not many readily identifiable points in a brain image.
3) Each landmark only defines 3 warping parameters (at most), whereas
   automated methods model many hundreds of parameters.
3) It is more labour intensive.

Methods based on manually identified point landmarks generally only model
coarse deformations, whereas automatic methods model more parameters. I
know there are a few papers that include comparisons between low and high
dimensional registration methods, but the only one I can think of relating
to detecting activations is:
JC Gee, DC Alsop & GK Aguire (1997). "Effect of Spatial
Normalization on Analysis of Functional Data".
Image Processing. Newport Beach, 22-28 Feb 1997
KM Hanson, eds. SPIE, Bellingham

There are of course a number of other approaches that involve registering
features that are not distinct points, but are surfaces, or extracted
features such as sulci and gyri. These methods may have some advantages
over volume based registration methods.

                    ---------------------------------------
In terms of volume based spatial normalisation on structural or
functional images, I'm not sure there is a reference to the literature
that relates to functional MR data, but there is one that relates to PET:
JH Meyer, RN Gunn, R Myers & PM Grasby (1999). "Assessment of Spatial
Normalisation of PET Ligand Images Using Ligand Specific Templates".
NeuroImage 9:545-553.

There are a number of issues to think about:
1) Spatial distortions in EPI data means that rigid registration with high-res
   structural images can not work particularly well for all brain regions.
   Some parts may coregister - but others wont. Misregistration is then
   propagated through to the spatial normalisation step.
2) Spatial normalisation in SPM99 only fits low frequency basis functions.
   Some parts of the brain require high frequency deformations to register
   them precisely, and so confound the estimation of the warps. In particular,
   variations in skull thickness and ventricular size have quite a strong
   influence. In most functional images, the scalp does not show up clearly,
   so there is less influence from variations in skull thickness. The
   brain-masking option is therefore not needed for spatial norm of these
   data.
3) The first step in spatially normalising an image is to spatially smooth
   it. Therefore, you dont really gain anything by using high resolution
   images for SPM99 spatial normalisation (although you possibly would if
   you were using some other spatial normalisation methods).
    

Best regards,
-John



%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%