Dear Greg,
The 3D DOF models are as follows:
6 DOF = rigid body = 3 rotations + 3 translations
7 DOF = "global rescale" = 3 rotations + 3 translations
+ 1 global scaling (all axes equally scaled)
9 DOF = "traditional" = 3 rotations + 3 translations
+ 3 scalings (each axis independently)
12 DOF = "full affine" = 3 rotations + 3 translations
+ 3 scalings + 3 skews (or shears)
I would never use 9 DOF - it is really there for historical
reasons (to emulate options available in other packages) but
I don't recommend it.
The 7 dof option is useful if you are trying to match scans
where you think that there might be significant changes in
scanner calibration or shim, but you don't want to let the
registration have the full 9 dof available to it. The reason
to avoid 9 dof is that for difficult registrations, such as
epi to anatomical, there is a tendency for it to use the scaling
to compensate for signal loss by stretching the A-P axis of
the epi to "cover" the areas of the frontal lobe where signal
loss has occurred. However, with 7 dof will let it partially
compensate for changes in volumetric calibration of the scanner
without allowing it to just muck about with the A-P axis like
this. If it tries to scale it to cover the frontal areas it
will also have to scale the S-I and L-R axes which will cause
the fit to be worse, in general, than doing little or no scaling.
You are right about the dofs for your epi - initial-highres -
main-highres registration, as long as you trust the volumetric
calibration between scans. Typically I will use 6 dof if the
scans have been taken in the same session, or within a relatively
short time frame (months) so that the scanner calibration, shimming
procedure, etc. has not changed. Otherwise I'd go with 7 dof
(or at least try it and see if it was better). For the main-highres
to standard you always want 12 dof.
Hope that clears things up for you.
All the best,
Mark
On 14 Dec 2006, at 19:23, Greg Burgess wrote:
> Hi FSLers,
>
> I have a set of quick and easy questions, to verify and clarify my
> understanding of registration. Can someone briefly explain what the
> degrees
> of freedom are in the FLIRT options within FEAT (i.e., 7dof and
> 9dof)? I've
> figured out that the affine transform can use 12 dof: 3D
> translation, 3D
> rotation, 3D anisotropic scaling, and 3D skew. What does the 7 dof
> mean, and
> why would I use it (what is the 1 extra dof, and why wouldn't I use
> 9 dof
> instead of 7 dof)?
>
> Lastly, I've read something in the mailing list that made a lot of
> sense (I
> think): "you should use 6 dof for epi to T1 registration. The
> assumption is
> that the brain did not grow or shrink between the time you acquired
> the two
> scans. Thus you can register with translation and rotation, but no
> scaling." However, this same assumption seems true for registering the
> initial high-res to the main high-res. Therefore, does this mean
> that I
> should use 6 dof (functional to init high res), 6 dof (initial high
> res to
> main high res), and 12 dof (main high res to standard)?
>
> Thanks in advance,
> Greg
|