Hi Marko,
Ah, you've found it. I haven't got around to thoroughly checking the
consequences, so it's currently set so that it doesn't iterate. The
scheme should be stable though
One of the reasons for this is that I figured I'd better let the TPMs
continue to have a strong influence on how the tissues are classified.
In the iterative scheme, they essentially act as Dirichlet priors for
the model. Without the prior, the algorithm would have the freedom to
simply find the best model of the data, under the constraint that there
are six tissue classes. This is not really what we are after, because I
strongly suspect that the six classes it finds would not be the ones
that most people want the segmentation to produce.
Another reason is that at the moment, the images are segmented
independently of each other. It is often the case that investigators
scrape together a few scans after they have finished segmenting the
others. In the current framework, those scans would be treated like all
the others.
For more insights on the kind of behaviour that may be expected for a
similar model (with skull-stripped data fitted to three classes), you
may wish to take a look at eg:
@article{bhatia2007groupwise,
title={{Groupwise Combined Segmentation and Registration for Atlas Construction}},
author={Bhatia, K.K. and Aljabar, P. and Boardman, J.P. and Srinivasan, L. and Murgasova, M. and Counsell, S.J. and Rutherford, M.A. and Hajnal, JV and Edwards, A.D. and Rueckert, D.},
journal={Lecture Notes in Computer Science},
volume={4791},
pages={532},
year={2007},
publisher={Springer}
}
Best regards,
-John
On Thu, 2009-08-27 at 14:42 +0200, Marko Wilke wrote:
> Dear All,
>
> when going through the code of spm_preproc_run, I noticed that what
> looks like a Dartel'ish approach is invoked (in line 50 and others) if
> the number of iterations is set to > 1, which, however, is not the case
> by default (i.e., it is 1, in line 47). I set it to 5 and it does run
> several times longer, as expected, but I have not systematically checked
> whether results are really worth the extra effort.
>
> So my question is: has anybody experimented with this setting? Or,
> alternatively, was there a reason to disable this iterative approach?
> Any insight is appreciated.
>
> Thanks in advance,
> Marko
--
John Ashburner <[log in to unmask]>
|