Hi,
what you could try is the following:
i) affinely register the individual FA maps to a template
ii) apply 'midtrans' to the 3 *.mat-files produced by flirt
iii) get the inverse using convert_xfm
iv) apply that to the template to get a midspace FA
You would then have to re-register the individual FAs and potentially average these prior to fnirting (to get a uniform non-linear reg across the 3 time points).
However, having said that I think even theoretically we have no "optimal" way or uniform gold standard to register longitudinal data yet. It all depends on the assumptions you are willing to make and the data you have. For example, if people were repositioned distortions may vary quite a bit. Then the best processing depends on the distortion correction and how good it works. Even signal loss can vary, e.g. you will observe less signal loss from adjacent sinuses if a subject comes in with sinusitis at one time point.
Essentially, you need to decide whether affine registration of the individual time points to each other for a given subject is really the way to go and at how many DOFs. But Steve has already indicated that the FMRIB crew is going into that.
Any comments, Steve / MJ / Jesper?
Cheers-
Andreas
________________________________________
Von: FSL - FMRIB's Software Library [[log in to unmask]] im Auftrag von Ryan Muetzel [[log in to unmask]]
Gesendet: Dienstag, 25. Mai 2010 22:05
An: [log in to unmask]
Betreff: [FSL] 3+ time point longidtudinal data
Hello,
We are conducting a longitudinal study and are using TBSS to warp each FA
map to standard space. With our DTI data, we've found slightly more
intra-subject variability in the nonlinear warp than we were expecting.
In an attempt to address this variability, prior to running tbss_2, we
align the time-1 and time-2 data into a half-way space (just as is done in
SIENA). From there, we compute an average FA map for each subject (time-1 +
time2/2) and feed that image into tbss_2 to determine the nonlinear warp.
We can then apply this warp to the individual time-point images using
tbss_non_FA. This method seems to reduce the intra-subject variability by
using the same nonlinear warp for each time point.
My question then is, is there a way to find the "mid point" among 3+ scans,
just as is done in SIENA with 2 scans?
Best,
Ryan
|