hello,
Van Leemput in his EMS toolbox starts with the M step and the classification
maps are initialized with the Atlas priors. Therefore there is no need to
initialize the parameters, they are computed as Ged says as weighted average
from the atlas probability maps. From my experience it seems to work pretty
well.
I did not pay attention to this part in SPM and I assumed it was the same.
Have many people experienced convergence problem with actual datasets in
SPM?
Olivier
__________________________________________________
Olivier Salvado, PhD
Case Western Reserve University | Case Center for Imaging Research
University Hospitals of Cleveland | Department of Radiology | Wearn B49
11100 Euclid Av. | Cleveland, OH 44106
Ph. (216) 983 3426 | Fax: (216) 844 4987
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On
Behalf Of Ged Ridgway
Sent: Friday, February 16, 2007 10:56
To: [log in to unmask]
Subject: Re: [SPM] Choice of starting values for mixture model estimation
Hi Chiara,
In SPM5 they are randomly initialised; from the Unified Segmentatin
paper: "Starting estimates for the cluster parameters are randomly
assigned", and from the code, see spm_preproc, lines 135-139.
In SPM2, my understanding of the paper "Multimodal Image Coregistration and
Partitioning-A Unified Framework" is that they are initialised as the means
of the voxels weighted by the priors, since the priors have been affine
aligned with the images already.
Presumably, in SPM5, the same idea could be used, if the initial
registration of TPMs to subject is adequate for this. Maybe John has tried
that already?
Best,
Ged.
Nosarti, Chiara wrote:
> Dear experts
>
> could anyone tell me how the starting values are chosen for mixture
> model estimation using the EM algorithm, as I have had a comment that
> this type of estimation is prone to local solutions. A reference I
> could look at this in more details would be also useful.
>
> best regards
>
> Chiara Nosarti
>
> __________________________________
> Chiara Nosarti PhD, Lecturer in Cognitive Neuropsychology, Division of
> Psychological Medicine, Room M5.01.05 Section of General Psychiatry,
> PO Box 63, Institute of Psychiatry
> 16 De Crespigny Park, Denmark Hill,
> London SE5 8AF, UK
> Tel: 0207 848 0133
> Fax: 0207 701 9044
>
>
>
|