Hi Jesper,
I mostly agree with you but there are two points that I think
sometimes get lost in registration/normalization discussions:
1. No two brains are the same, anatomically or functionally.
Registration implicitly assumes the problem is purely topological --
namely that there is exists some perfect transformation that will make
brain A exactly the same as Brain B (where B is normally a
template). In fact, I'd suggest this same assumption is implicit in
your comment:
> If we assume that your registration algorithm is unbiased that
> position will get closer and closer to the "true" position as your n
> increases. But the "better" your registration algorithm is the
> faster it will converge to the "true" position (with "faster" I mean
> for smaller n). Hence, the better your registration algorithm is the
> closer to the "true" locations will your reported coordinates be.
You carefully put "true" in quotes but my arguments is that there is
no such thing. If that is correct -- and anatomists have been telling
us that for nearly 100 years -- then the value of registration to a
common space is to make data from different subject analyzable in an
unbiased fashion but not to suggest that whatever standard space being
used is necessarily meaningful. In fact, that is why Russ and I
argued that group results should be displayed on the group mean brain,
because that accurately represents some truth (for the subjects that
were tested, at least) as well as the true anatomical variability in
the data.
> it is true that the registration algorithm have an impact on the
> results. If not people wouldn't bother keep coming up with new and
> better algorithms.
It definitely affects the results but they are not necessarily
"better" because "better" is extremely difficult to define. If there
is no "correct" transformation -- even theoretically -- then it
becomes critical to define what our criteria are and I would suggest,
there should be some broad agreement in the field on these criteria
otherwise one can always invent a new algorithm and claim it is better
on some measure.
So my point was simply that one criteria for "better" is reproducible
across labs, scanners, software packages, etc. in part because more
and more groups are using meta-analyses or just big analyses of data
archived in large storage systems when in fact, the preprocessing
steps may be artificially introducing unwanted noise.
This is not the only criteria -- precise functional-anatomical studies
in individuals would not find this a useful definition at all, but
then they are not using templates either. As long as a template is
being used, the aim must be to combine data into a standard space (of
some sort), and often the additional aim is that the results of these
analyses can be compared to studies done elsewhere.
> That means that if you want to compare reported coordinates between
> studies (assuming here we are talking about identical paradigms)
> they will be more similar the better your registration algorithm is.
Only if there is an underlying "truth." If not, it's just noise.
2. With respect to spaces, I think we completely agree -- similar
naming conditions do not mean identical spaces. In other words, there
is no such thing as "MNI space." In fact, there are several different
MNI spaces that are all similar in some respects, but not identical,
and they are defined by the specific template being used. So MNI152
is not the same as MNI305 or MNI452, despite being from the same lab
and based on similar underlying data and procedures. I think this was
the main point that Van Essen and Dierker were making in their paper.
BW, I realize I am swimming against the neuroimaging-tide here but I
think that is at least partly due to the community using the tools
without critically considering these issues. The mathematicians who
develop the tools are providing an incredible service to the community
and it is our responsibility as users to provide feedback in kind to
help carefully define the problem. I wonder how many people stop and
think what it means to use a linear vs. nonlinear registration or a
given template and how those choices affect their results. So I find
discussions like this really beneficial and would like to thank Monas
for raising the issue and Jesper for his excellent comments.
Hopefully others will join in and add their perspectives?
Joe
|