Print

Print
Hi David,

You are correct that you could apply SMM to other images, e.g. copes. The best thing to do would be to have an SMM on the regression parameters (or linear combinations of them) in the GLM as part of an all-in-one model - this would fit the GLM at the same time as classifying and spatially regularising the classification. However, this is time consuming. So the closest to this, whilst being computationally practical, is to apply the mixture modelling (after the GLM fitting) to the normalised copes, i.e. to the z-stat images.

You are correct that the documentation is a little sparse on the output of SMM! This is something we can easily rectify. But your deduction that w1=nonactivation, w2=activation and w3=deactivation is correct. These are the key files in the output that tell you about the classification. The mu_mean, var_mean files tell you the posterior mean estimates of the mean and variance respectively of the class distributions.

The ability for mixture modelling to adapt the non-activation/null distribution to deal with shifts or changes in its widths due to GLM modelling inaccuracies is one of the attractions of mixture modelling (along with inference flexibilty). Instead of asking the question “Is the activation zero or not?,” we can ask the question “Is the activation bigger than the overall background level of ‘activation’?.”  However, this all only works well when the distributional assumptions are well met, in particular that the non-activation/null distribution is well modelled by a Gaussian. When the modelling inaccuracies particularly due to structured noise artefacts (e.g. "resting" state networks) get severe enough, this assumption can go awry and so this is perhaps the best reason as to why SMM or nonspatial MM is not finding wider use. Our plan is that future developments in structured noise modelling, as part of the GLM modelling, will address this problem. In the meantime I would recommend that when using SMM or MM you should be on alert for situations when it looks like this assumption is being violated.

Hope this helps.

Cheers, Mark.

----
Dr Mark Woolrich
EPSRC Advanced Research Fellow University Research Lecturer

Oxford University Centre for Functional MRI of the Brain (FMRIB),
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK.

Tel: (+44)1865-222782 Homepage: http://www.fmrib.ox.ac.uk/~woolrich

On 16 Feb 2007, at 12:42, Glahn, David C wrote:

Spatial Mixture Modeling (SMM)

I have a set of questions concerning the spatial mixture modeling application (mm).  

First, while I used mm with a set of zstat images and generally got good results, it occurs to me that the SMM procedures may work better on images that have not been normalized into a Gaussian distribution.  Is it best to apply SMM to zstat images or some other type of image (% signal change, cope)?

Second, is their any documentation explaining the output files from the mm utility?  With trail and error, I was able to determine that the W2_mean image is the “activation” image and the W3_mean image is the “deactivation” image in the html output.  Is this correct? What are the other images and are they useful in interpreting statistical parametric maps?

Third, the results from 1 of the 6 images analyzed with mixture models differ dramatically from results obtained with the cluster algorithm.  Specifically, the mm activation map includes large activations not included in the other.  Upon inspection, I discovered that the null distribution from this image is wide and somewhat defuse.  In such a case, how would one interoperate mm results?

Best,

David