> (1) When warping to MNI space with a low-resolution image, you get a
> striped pattern that is co-planar with the acquisition plan. Between
> the stripes the values of the voxels are 0. This was confirmed by
> zooming in and using NN interpolation in the checkreg function. This
> pattern does not exist when warped to the DARTEL template. Why does
> warping to MNI space lead to 0 value voxels within the brain? This
> seems to be very problematic for processing data.
This option is usually used for anatomical scans, which have a higher
resolution. My aim was to exactly preserve the volumes of tissue, so
instead of sampling the image using the inverse transform, and then
scaling according to the Jacobian determinant, the approach here pushes
the voxels to their new locations. It can lead to slight aliasing
effects for anatomical scans, but these are usually dealt with by the
smoothing. For low resolution images, the aliasing effects are more
extreme, which is what you are seeing.
>
>
> (2) Using high-resolution data, you get a dotted pattern in the MNI
> space data, which also seems to be non-optimal -- not shown.
Slight aliasing is to be expected, but because the images are blurred,
then it is not usually apparent.
>
>
> (3) When warping to MNI space and not modulating, the data gets
> smoothed into the non-exisitng space. As has been noted previously,
> the data in the brain seems to be unaffected and that its only the
> edge that has this problem. On several other images, whenever there is
> incomplete coverage of the brain/skull where the brain/skull ends the
> image gets smoothed in a weird manner similar to that in the attached
> pictures. This only occurs when the images are warped to MNI and only
> when they are not modulated. What would cause the weird smoothing in
> only this processing stream and not the modulated images?
The deformations from DARTEL are more extreme then those from the older
spatial normalisation. For this reason, there is more chance that some
voxels become extremely stretched, and hence contribute a
disproportionate amount of signal to the smoothed data. The solution
was to weight the smoothing, so that voxels contributed a more equal
amount of signal. There is more said about this in the Manual. See
some of Christian Buchel's emails on the list about his masking
solution.
https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0908&L=SPM&P=R895&X=0BB9794F94B0034DAA&m=3254
If you ignore the stuff outside the field of view of the warped images,
you will see that the intensities within the field of view are better
behaved for the images in the 2nd row. For images in the bottom row,
you'll find that the intensities towards the top and bottom begin to
taper off towards zero, within the part of the image that you intend to
analyse. Smoothing of images requires some assumptions about what to do
at the boundary of the field of view. In the bottom row, the assumption
is that the value is zero outside, which contaminates more of the signal
inside. This will lead to further systematic differences arising
through variability in head positioning. I would expect the images in
the middle row to give a larger coverage of usable signal than those
along the bottom. The only issue is that you'll need to provide a mask
when you fit the GLM.
Best regards,
-John
--
John Ashburner <[log in to unmask]>
|