Dear Joerg,
When you say the registration "doesn't work" or there is an "error" - do
you mean that it is slightly shifted in z (by about 3-4mm) which is
actually
a sub-voxel shift with respect to the slice thickness of ref_1H?
That is all the difference that I can see. Assuming that this is the
problem,
to get an even more accurate (sub-voxel) alignment I have used an
input weighting
volume to exclude the edge of the brain as the BET results are not
perfect.
Here is what I did (based on the images generated from the previous
bet results
and using the previous alignment as a very close starting position):
avwmaths ref_1H_bet -thr 0.01 -bin ref_1H_bet_mask
avwmaths ref_1H_bet_mask -ero -ero2 -ero2 -ero2 -ero2 -ero2 -ero2
ref_1H_bet_mask_ero
flirt -in ref_1H -ref anatomy_bet -inweight ref_1H_bet_mask_ero -init
ref2anat.mat -nosearch -dof 6 -out ref2anat2 -omat ref2anat2.mat
Note that the erosion in avwmaths is done to effectively erode about
5mm in
the z-direction (the -ero) and 6mm in the x-y directions (the -ero
plus the
other -ero2 statements) as otherwise there is either too much erosion
in z, or
not enough in x and y to eliminate the problematic edges from the BET
result.
The resulting registration I believe is even better than the one you
can get
from using mutualinfo and the search restrictions you mentioned
before. Plus
it has the advantage that it is more systematic (as restricting the
search in this
way is probably just biasing it to find a local minimum that you like).
As a check, here is the matrix it gave me:
0.999945 -0.00968041 -0.00401788 -36.8605
0.00851318 0.973763 -0.227409 -5.23363
0.00611387 0.227362 0.973791 101.652
0 0 0 1
I really do strongly recommend using BET, as your ref_1H image shows
very
bad partial volume effects at the top of the head (due to the very
thick slices)
and this just cannot match well with the higher-res anatomical image,
and so it
will bias the registration of the brain structures, which is not what
you want.
In addition, when looking at the results be careful to allow for the
fact
that the ref_1H volume will be interpolated to higher res, but still
has thick
slices for the input, and so this will tend to stretch some
structures in the z
direction - especially very high contrast ones such as the
ventricles. In general,
when looking at T2-weighted images vs T1-weighted images, then
ventricles appear
larger in the T2-wt to the eye, since it is much more sensitive to
small amounts
of partial volume of CSF, hence higher slices of the ventricles, in
particular, can
seem mismatched, although if you look carefully at the T1-wt image
you can see
the subtle intensity change there. The best way that I have found to
check for fine
accuracy is to look at the sulcal patterns on the axial slices one at
a time,
especially near the top of the brain. I found that in this case the
sulcal
patterns matched very well - again, keeping in mind that it will look
like you have
more CSF in the T2-wt images.
Let me know if you get good results with the commands I used above.
All the best,
Mark
On 17 Mar 2007, at 08:19, Joerg Magerkurth wrote:
> Hi Mark,
>
> sorry, it didn't work. I have tried it with your parameters, also
> with bet2
> and dof=7. flirt version is 5.4.2. The main error in fslview is a
> shift in
> the z direction. Have no other ideas than using flirt without bet,
> this
> looks better but not perfect. What's criticaly when I don't using bet?
>
> If it helps to find the problem, here is the matrix from flirt.
> Check it
> with yours.
>
> 0.99996 0.00767671 -0.00456802 -39.9283
> -0.00870712 0.951869 -0.30638 -0.400576
> 0.00199617 0.306408 0.951898 86.8124
> 0 0 0 1
>
> Best
> Jörg
|