Dear Antonio,
> I have some basic questions regarding FLIRT. I must say from the onset, and
> this will be evident in the next lines, that I only understand the gist of
> how it works (from the slide show on the FMRIB site and your chapter on the
> ?Medical Image Registration? J.V. Hajnal et al. book).
>
> 1. Will an EPI to standard brain registration benefit from an intermediate
> hi-res T1-W scan? I am guessing it can only do good, but wonder about the
> accumulated error of doing two registrations instead of one.
Firstly, EPIs tend to have poor resolution and poor contrast.
Furthermore, standard brains are often averaged over many individuals
are are somewhat diffuse. So, registering these two images directly is
very error prone.
If, instead, you have a highres scan (doesn't have to be T1) of the same
individual then this improves matters *a lot*. The reason is that the
registration from the EPI to the highres only needs 6 or 7 DOF as the
anatomy as the same. That is, the registration does not have to adjust
the overall shape of the brain (besides some potential adjustment for
scanner changes). As this is a more constrained registration, and the
highres image has better resolution and contrast, the quality is
significantly better. Then, the second stage is to register the highres
to the standard. Again the improved resolution and contrast of the
highres make this registration better, even though it now requires the
full 12 DOF to adjust for changes in overall shape of the head/brain
between the individual and the averaged standard. Given the two good
quality registrations, the combined version is also good quality and any
amount of accumulated error is totally negligible in this context.
> 2. Would a particular EPI contrast make FLIRT happier? I plan to use it for
> registering short-TE EPI volumes, on which there is very poor CSF contrast.
Many factors affect the registration. Contrast is one, but so is field
of view, artefacts (e.g. B0 inhomog.), and of course, resolution.
Therefore there is no general answer to this question. However, if you
have almost no contrast between WM/GM/CSF (like a pure PD) then all the
registration has to work with is the outer surface of the brain/head
(tissue to background contrast). This can cause problems, but there is
also a good chance that it will work fine. You can only try...
> 3. I understand that FLIRT registers two brains using their intensity values.
> The end result should be a brain that is aligned and scaled to the reference
> brain through a 12DOF affine transformation. Such registration then creates a
> common frame of reference for both brains (and I write all this just to
> corroborate...).
Yep, that's right.
> But, I am confused by the following: When I transform an EPI brain
> directly to the standard brain, I notice that as I surf through the axial
> slices of the registered brain, the perimeter of the cortex is similar but
> not precisely that of the standard brain. Is this discrepancy simply
> outlining ?residual? differences between the subject?s brain and the standard
> brain after registration, or is it due to errors in the registration itself?
Firstly, don't do the registration directly - use a highres intermediate
(see answer 1 above). As to exact matching - it is an affine
transformation - only a general non-linear registration scheme will
achieve an exact match as there will be small differences between every
individual and the standard which cannot be corrected in the more
limited framework of affine registration. To get an exact match you
will need a non-linear registration method.
Best wishes,
Mark
|