> We have seen the exact same thing, kind of like high vacuum was applied to
> the top of the head. Actually worse than the fact that this occurs, is that
> it occurs differently for each subject so group normalization sucks. This
> seems to be due to signal inhomogeneity from the coil, the sequence or
> whatever. We had the same problem with the Transmit/Receive and the
> 8-channel head coil. Our physicist hasn't had much luck with adjusting the
> MPRAGE sequence per se (but see below) and we were unable to implement
> satisfactorily the MDEFT sequence written up by Ralf Deichmann. By the way
> even though the MPRAGE images don't "look" that bad, if you do a bias
> correction on one of the bad volumes and then plot the bias field the
> region requiring warping becomes obvious.
Spatial normalisation in SPM2 doesn't include any model for intensity
nonuniformity - so if there is a lot of this artifact in your images, then it
can lead to problems. The segment button in SPM5b includes bias correction
within the warping model, so should be able to handle this sort of thing
rather better.
Another possibility (of many) is that the spatial normalisation parameters
are being estimated from resliced images. Resliced images often have bits
chopped out of them, which the spatial normalisation does not deal with
especially well.
>
> Anyway I've tried a number of techniques to fix this. As you note
> segmenting the brain and normalizing the gray matter segment to the apriori
> gray matter image can help, but doesn't always solve the problem because
> the gray matter segmentation is also affected by the inhomogeneity. So the
> gray matter at the top of the head is subtly deficient and then normalizes
> poorly.
The problem here is that the initial affine registration does not model
intensity nonuniformity (bias). With sub-standard affine registration, then
the tissue probability maps (used as priors for the segmentation) won't be
accurately overlayed - leading to sub-optimal segmentation. The actual
tissue classification model of SPM2 usually deals with bias pretty well (far
better than for SPM99).
>
> I spent one weekend coming up with a complicated sequence of steps to
> minimize the distortions, which required adjusting the defaults, bias
> correction, skull stripping the brain, segmenting and then normalizing.
> I'll be happy to send the steps to you if you want, but the bottom line was
> it didn't work all the time.
I spent many weeks working on this and came up with the solution in SPM5.
There is still scope for improvement, but it is definately a step in the
right direction.
Best regards,
-John
|