Thank you for trying those out.
For "2007" I was being a bit picky (it's okay, but SPM normally does a
better job). Some grey and white matter segmentations have leaked into
dura and superiorly a little true grey matter had been lost, CSF has
expanded more than it normally does. I thought it looked like it may
be an initial misregistration, but could also be intensity based
(having now tried manually aligning first with the same result.)
"1775" I'll have to confess these images have been stored in a way
which removes the exact scanner coordinates (gross axes are correct),
it looks like that was originally a slightly oblique acquisition.
Difficult to get it restored for all of them unfortunately, this would
also explain the poor z-offset too. The original is not quite so
tilted. ADNI quality control will have generally been more worried
about motion or noise in any case. Overall it's a small percentage
that have gone wrong, I'm fairly happy with the idea of running
co-registration beforehand, possibly on just those that fail the first
time actually, to avoid messing with things too much.
Best wishes,
Ian
On 30 April 2015 at 19:03, H. Nebl <[log in to unmask]> wrote:
> Dear Ian,
>
> I've just looked at the two files with SPM12, for "2007" the segmentation seems to work well without any adjustments. The displacement along z of the raw image is already large though, about 20 to 25 mm relative to the MNI templates, thus other routines might indeed fail depending on settings/penalities. An extra coregistration step before the segmentation does correct the displacement nicely, also with regard to the rotations applied.
>
> Concerning "1775", the origin is shifted to some extent, but that subject is especially tilted backward when looking from the side, maybe 20° relative to the templates. This is the case if the head was not placed "high enough" on cushions, then it "pitches backward". Sometimes subjects have neck problems making it uncomfortable with the pads below their head, thus they have to be removed. Sometimes the usual pads are just not enough (maybe in overweight people). Unfortunately the default coregistration and the coreg step within the segmentation are often not capable to correct this deviation as it's an atypical positioning / the required parameters along pitch are probably too large and the algorithm ends in a local maximum/minimum instead. auto_reorient.m also fails in that case.
>
> In general, there might always be some instances in which algorithms fail, as those available are rather simple. To stay on the safe side you could turn to a manual reorientation according to AC, AC-PC, providing an equally good/bad starting point for all your images, but this is of course very time-consuming for large-scale studies. Alternatively and depending on the raw data, you might decide to add e.g. an initial coreg step, then check the data after segmentation and discard the bad ones. And concerning the quality check from the ADNI, I guess these were rather basic aspects and not whether an image can properly be aligned on a template.
>
> Best,
>
> Helmut
>
--
imalone
|