In principle, such prior knowledge could be helpful, but the code would need some slight modifications in order to include such knowledge. Doing it optimally would need slightly more than just a few previously estimated bias fields, as the estimates of each of these is influenced by the priors that were used for estimating them.
The whole thing gets a bit circular, and would need an iterative algorithm that loops over all images in order to properly solve the problem. Similar reasoning also applies to estimating deformations for nonlinear inter-subject alignment, and also for generating many of the priors that would help segmentation algorithms (such as priors on the intensity distributions of the various tissue classes, and also the tissue probability maps that are used as spatial priors).
My prediction is that such hierarchical Bayesian approaches will become more widely applied in brain image processing applications. Currently, such machine learning approaches would be pretty fertile ground for long-term research - especially as all the "low hanging branches" have now been picked pretty bare.
Best regards,
-John
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Marko Wilke
Sent: Monday, February 26, 2007 9:07 AM
To: [log in to unmask]
Subject: [SPM] Bias field
Dear All,
I stumbled upon the following passage in the segmentation help text:
'A more accurate estimate of a bias field can be obtained by including
prior knowledge about the distribution of the fields likely to be
encountered by the correction algorithm.'
I wonder if this only refers to the amount of inhomogeneity (likely a
ballpark number) or it is it possible to, for example, use another
algorithms output as a more informed starting estimate. Alternatively,
one could take the bias field from a first iteration and feed that into
a second pass segmentation run. I realize this would take more time but
for very inhomogeneous data (surface coils, high-field etc.) it may
result in an improvement of tissue labeling.
So bottomline, my question is: Has anyone experimented with providing 3D
inhomogeneity estimates as starting parameters for the spm5 segmentation
algorithm?
Best,
Marko
--
=====================================================================
Marko Wilke (Dr.med./M.D.)
[log in to unmask]
Universitäts-Kinderklinik University Children's Hospital
Abt. III (Neuropädiatrie) Dept. III (Pediatric neurology)
Hoppe-Seyler-Str. 1, D - 72076 Tübingen
Tel.: (+49) 07071 29-83416 Fax: (+49) 07071 29-5473
=====================================================================
|