> I have some follow up questions:
> > Assuming that the warping is exact (a big assumption), then the warped
> > > grey matter of one image will be identical to the unwarped grey matter
> > of > the other. The expression is essentially i1.*i2 - i1, which is the
> > > modulated image (theoretically identical to a modulated warped
> > other > image) minus the unmodulated image.
>
> So, when you subtract the unmodulated GM image from the modulated one, what
> is the result you are left with? Would you say it is the image of GM
> volume change over time or something like that?
Yes.
> This raises another
> question: if we are doing stats on these images from a bunch of people,
> the time interval between the "early" and "late" scans will differ with
> each individual. Is there a better way to model the interscan interval?
You could assume that the volumes change linearly with time. This is highly
unlikely (as continuing to decrease linearly would eventually result in
negative volumes), but for small differences it could be a reasonable
approximation.
> It seems to me this method assumes the time between scans is the same for
> all people in the analysis.
The relevant columns in the design matrix could be scaled according to the
time difference. A polynomial expansion of time difference could also be
used in order to model nonlinear effects.
>
> > > Also, I see your point about smoothing log(det), but there are also
> > > statistical reasons why people want to take logs; do you think it
> > > would be okay to smooth first and then take logs?
> >
> > You hit problems in regions containing only zeros.
>
> Would it be feasible to do log(det+eps) or something similar to fix the
> zeros problem?
I don't know what is best here.
Best regards,
-John
|