The first part of the segmentation is an affine registration with the tissue
probability maps, using Mutual Information as the objective function.
Although I incorporated bias correction in the nonlinear warping part, I
didn't get around to including any for the initial affine registration (which
was never written up as a paper). This means that any bias is not modelled
at this stage, and so leads to problems of the kind you are encountering.
The proper solution would be to have affine registration and bias correction
combined. As an ad hoc workaround, I would suggest either:
1) Disable the affine registration (but give slightly better starting
estimates by reorienting the data).
2) Try a 2 pass procedure.
Best regards,
-John
> I have been playing with the unified segmentation/bias
> correction/normalization in SPM5 and I have a question: should
> segmentation and normalization be performed on bias corrected images? In
> other words, run the "segment" procedure once to produce a bias
> corrected image, and run it again using the bias corrected image to
> produce the segments.
>
> Here is an example that illustrates why this seems like a good idea
> (compare images 2 and 4):
> 1 - raw image (mean of motion-corrected series)
>
> 2 - segmented gray matter (output from applying Segment to image 1)
>
> "bad"
>
> 3 - bias-corrected mean (output from applying Segment to image 1)
>
> 4 - segmented gray matter from bias-corrected (output from applying
> Segment to image 3)
>
> "good"
>
>
>
> The bias correction seems to do a great job, and the second-pass
> segmented gray matter image (#4 in table) seems good.
>
> However, I had assumed that the unified procedure avoids the need for
> iterating through these preprocessing steps, but that assumption seems
> incorrect. I thought that the unified procedure conceptually applies
> the bias correction and then segments according to bias-corrected image
> intensities (and priors). However, in this albeit extreme case this does
> not seem to be what is going on.
>
> Regards,
> Paul Macey
> Assistant Researcher
> Department of Neurobiology, UCLA
> 10833 Le Conte Avenue
> Los Angeles, CA 90095-1763
> Email [log in to unmask]
|