[Insight-users] Re: itkExpectationMaximizationMixtureModelEstimator

[Christoph Heinzl]

Hi all,

I plugged in my image data into the
itkExpectationMaximizationMixtureModelEstimator using an ImageToListAdaptor.
I'm still facing the problem that the estimator stops after the first
iteration, even if I use roughly the expected resulting parameters as
initial parameters. I built up my application similarly to the example in
the software guide (10.4.4 Expectation Maximization Mixture Model
Estimation) : 2 GaussianMixtureModelComponents plugged into the estimator.
Has anyone experience with itkExpectationMaximizationMixtureModelEstimator?
Why does the algorithm stop?

By the way: the KMeans estimator works fine on the data. So I classified my
dataset into two subclasses and calculated mean and covariance using
MeanCalculator and CovarianceCalculator which worked out fine.

Thanx for your help!!
Chris


On Nov 22, 2007 10:31 AM, Christoph Heinzl <c.heinzl at gmail.com> wrote:

> Hi
>
> In my recent application I need to fit gaussian normal distributions
> into the histograms of my volumetric CT datasets. Usually there are
> two or three different classes in my data: air regions, material1,
> material2. So, first of all I would like to identify the means of my
> clusters, which is working fine with K-means algorithm. In order to
> fit now the normal distributions, I some kind of expectation algorithm
> should work fine. As I saw in the ITK book and from prior postings
> there is the itkExpectationMaximizationMixtureModelEstimator which
> should be able to find both the means of the clusters and the variance
> of these distributions. But my question is: How should the input data
> for this funtion look like? If I input an array containing my
> histogram data (from bin 0 to 65535) I don't get a valuable result.
> The algorithm stops at the 1st iteration returning my preset initial
> values for the clusters. Should the input data be presorted somehow?
> Is the raw data of the dataset expected als input or the cumulated
> histogram information?
>
> THX in advance,
> Chris
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://public.kitware.com/pipermail/insight-users/attachments/20071128/de46fbe0/attachment.html