Dear Jesper,
thanks for your comment. What we did was to segment the images using the
spm99b
routine, choosing the option saying that the target images are
non-normalized.
>From my reading of the code, I think that the program then proceeds to map
the
a-priori template images into the target MRI space before using them, so I
hope
your concern does not apply. Is this correct? The images certainly look fine
(in
fact, they look great) and accurately represent the grey-white matter
distribution of the target image.
We'll follow up on Andrew's suggestion and report back later.
Thanks again and best regards, Andreas
-----Original Message-----
From: Jesper Andersson
To: Meyer-Lindenberg, Andreas (NIMH);
Cc: Japee, Shruti (NIMH-CBDB)
Sent: 1/10/2000 4:38 AM
Subject: RE: Questions about affine normalization of structural MRI data
(lengthy)
Dear Andreas,
>Dear Andrew and list,
>
>many thanks for the previous reply which has been very helpful and
enabled
us to
>proceed with our project. Unfortunately we'll need some further
guidance
as we
>have not been able to validate our basic assumptions.
>We have used an affine transform (only) to map gray and white
segmentation
maps
>into Talairach space, resampling from 0.9375x0.9375x1.5mm voxel size
(voxel
>volume 1.32 mm3) to 1mm isotropic. Sinc interpolation was used. The
>Transformation matrix Q was
>
> X1 = 1.201*X - 0.137*Y - 0.073*Z + 1.270
> Y1 = 0.165*X + 1.051*Y + 0.444*Z + 1.284
> Z1 = 0.004*X - 0.514*Y + 1.113*Z - 2.366
>
>(from mm3 to mm3) with an absolute determinant of 1.71, which should be
equal to
>the ratio of voxel volumes before and after the transformation, as we
discussed
>previously. However, we get (counting voxels greater than 0):
>
>gray: 797705 voxels = 1051662 mm^3
>ngray: 2094698 voxels = 2094698 mm^3
>
>Volume ratio: 1.992
>
>white: 476199 voxels = 627801 mm^3
>nwhite: 1607171 voxels = 1607171 mm^3
>Volume ratio 2.559
>
>So both ratios are quite a lot higher than the expected 1.71, but even
more
>disconcertingly, the ratios are not the same for the two maps (as they
should be
>from my understanding of affine geometry). The images look fine. We
have also
>counted voxels above a certain threshold > 0 (0.05, 0.1) with the same
result.
>The options for masking of brain and target volume also have no effect
on
this.
Not having read your previous mail I don't know if you use SPM for the
segmentation as well, but if you do there is a possible explanation. The
segmentation in SPM assigns each voxel to either grey, white, CSF of
extracranial depending on their image intensity values (data driven) AND
the a priori probability of finding that particlar tissue type at that
particular location. The latter probability is assessed through a
probability map, in Talairach space, based on a large number of normal
brains. When you perform the segmentation on the data prior to the
affine
normalisation the prior estimates are going to be wastly wrong, in
particular since your transformation matrix involves quite a bit of
rotation in addition to a ~20% scaling in each direction. Hence, I think
your "gray" and "white" values above are quite wrong. You should be able
to
check this by looking at your segmented maps and comparing them to the
original images.
>
>We are quite puzzled by this result and would greatly appreciate your
help. Best
>regards and a nice start of the week,
>Andreas
>
Good luck Jesper
Jesper Andersson
Wellcome Dept. of Cognitive Neurology
12 Queen Square
London WC1N 3BG
phone: 44 171 833 7484
fax: 44 171 813 1420
|