| I have been working with the spatial normalization routines in
| spm99 and have run into a few difficulties. First, I realigned
| all the data sets creating a mean image (we are working with 145
| mm coverage 3.75 x 3.75 x 5.00 resolution x 200+ scans) for each
| session. Then I tried to normalize (using all the default
| options) these mean images to the EPI template provided with
| spm99. I ran about 100 subjects last night and in about 25% of
| them the normalization produced a poor fit. The problem is
| largely in front of the AC line in that the front of the brain is
| stretched out much farther than it should be (I can email
| postscript output if anyone wants to see). I then played around
| (as the online help suggests) trying to re-orient the mean images
| (after deleting the .mat files of course) such that they matched
| (visually) the template's orientation to hopefully provide the
| algorithm with a better starting estimate. At most the data
| needed to be rotated about .3 degrees. This did not help. Then I
| looked more closely at all the mean images in which there were
| difficulties and realized that most of them were slightly bigger
| brains (in the y direction) than the template. I then tried to
| rescale them into smaller brain using the new display facility in
| spm99. This did not help. Then I tried to normalize them in
| spm96 (using the spm99 EPI template) and it worked great. Then
| I went ahead and plugged in the same defaults as spm96 into spm99
| (mainly 4.x5 x 4 nonlinear basis functions, instead of 7x8x7) and
| it still failed. Then I removed the 'masking' option and,
| still using 4x5x4 basis functions and it worked great. Every
| brain was properly normalized (as best I can tell with our
| display utilities). I am wondering if anyone else may have seen
| this before and whether I should just go ahead and proceed with
| stats. I would appreciate any comments the spm authors might
| have on this as well.
I would recommend disabling the brain masking option for spatially
normalising images with very little signal outside the brain. The
default settings are best for T1 weighted images, but may not be
optimal for EPI images, or even PET images. Brain masking is
intended for weighting the spatial normalisation so that it is
based only on information within the brain, and not on matching the
scalps. This is useful because variations in skull thickness can
not be modelled particularly well with only about 1000 smooth basis
functions
| While I am at it, I was curious about a couple of other
| functions. If I realign my data (origins set to AC) and create a
| mean image - and then decide that I would prefer to reorient them
| (b/c of this rotation we seem to have in all our data) - is there
| any reason that I can't re-orient the realigned images? My
| perusal of the code suggests that the re-orienting process first
| reads all the .mat files (from the realignment) and then applies
| the new orientation parameters. So, my guess is that it will not
| matter. Confirmation of this would save about 100 hours of cpu
| time. Thanks.
You can quite happily re-orient realigned images via the <Display>
button. The code pre-multiplies the existing .mat files with
a matrix derived from the specified rotations and translations.
This means that when the images are transformed, their relative
positions with respect to each other remain consistent.
| Next, I was thinking of changing the spatial normalization (to
| save disk space and time) so that it did not have to reslice the
| images until after smoothing had been applied. That is, can't one
| just take the spatial normalization parameters determined for the
| mean image and create new .mat files for each image in the
| session? Then modify the smoothing code to look for the .mat
| file and apply it before smoothing. Alternatively, perhaps just
| apply smoothing at the normalization step? Are there any reasons
| not to do this? It would save a lot of time and disk space.
| Thanks in advance.
I don't think this would work. The spatial normalization parameters
are stored in the _sn3d.mat file rather than just the .mat files,
and contain information about nonlinear warping rather than just an
affine transformation. The simpler solution for saving disk space
(but not time) would be to modify the smoothing code so that it
deletes the original file after it has been smoothed.
All the best,
-John
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|