Hi Jesper,
I'm afraid we may have to agree to disagree on this one. :)
The problem is not whether the images can be made to look more like
one another -- they can be. The issue is whether this is a meaningful
thing to do. Consider the case of V1 -- which is a region of the
occipital lobes that can be defined functional based on retinotopy.
Anatomically, it more or less sits in the calcarine sulcus but its
extent and specific topography varies from person to person which is
why if one wants to study it with fMRI in humans, it is necessary to
do a retinotopy scan. Any transformation that made subject 1's V1
match subject 2's V1 by forcing them into a common space isn't helpful
-- in fact, it is just the opposite because it is explicitly removing
information that is present in the data (i.e. that the two areas are
not the same).
And it's worth noting that the issue becomes more pronounced as one
moves away from primary sensory/motor areas. So called higher-order
association areas have greater anatomical and functional variability
than "earlier" areas.
Two final points that follow on from that:
1. Because the critical information is not present in the images
themselves, even "expert" manual analyses based on gross anatomical
features will never suffice to make two different brain the same. I
agree with you that loads of groups seem to be actively working along
these lines, but from my perspective -- as a neuroimager with an
interest in anatomy -- the endeavor itself is flawed.
2. Despite this variability that doesn't mean one should give up on
group studies (which is a position that some groups have advocated,
e.g. Fedorenko & Kanwisher (2009) Language and Linguistics Compass,
3(4): 839-865). It just means that if we want to make meaningful
group comparisons, particularly across studies, then we need to accept
this intrinsic variability as part of the interesting problem that we
are trying to solve. We don't want to either throw it away OR
artificially inflate it by introducing additional (unnecessary)
variability as a result of our registration procedures. Linear
registration meets both of these criteria; nonlinear registration
violates both.
Ok, thats it from me. Thanks again for a stimulating discussion of a
topic that receives less attention than it should.
All the best,
Joe
|