 |
 |
To try to answer your queries: 1) Many people consider manual segmentation to be the gold standard to which all segmentation algorithms should aspire. By this argument, visual inspection should be an appropriate strategy. 2) It is possible that reducing the number of Gaussians would help. One of the problems with mixture of Gaussians models - particularly when the data have a finite number of different discrete values - is that one of the Gaussians may fit a single intensity (i.e. the variance will approach zero). I don't know how many Gaussians would be optimal for your data. Try it to see what works best. In principle, a Bayesian model selection (or similar information theoretic procedure) should be able to select the best model. Given that the segmentation in SPM doesn't use this, then visual inspection will have to suffice. 3) A mask would only help in regions where the model does not explain the data well - e.g. where there are lesions. A better solution would be a more general model that could account for pathology. But again, SPM doesn't do this. 4) A poor characterization of non-brain intensity distributions could reduce the accuracy. Voxels that are not characterized as non-brain by the algorithm are considered to be brain. 5) If the characterization of the tissue class distributions is problematic, then the information from the priors is likely to be even more important than it would be otherwise. Both these components contribute to the classification. Best regards, -John -----Original Message----- From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Mehta, Sonya H Sent: Tuesday, June 05, 2007 11:53 PM To: [log in to unmask] Subject: Re: [SPM] SPM5 segmentation of skull stipped images Dear SPM experts, I have performed the preprocessing for a VBM analysis using SPM5 (Christian Gaser's vbm 5 toolbox) on MR data that did not include all of the skull. Furthermore, the amount of skull included in the scan varies from subject to subject. As such, I used skull stripped images as the input into SPM's unified segmentation algorithm. My understanding is that in using SPM5's unified segmentation, the tissue priors were used for both normalization and segmentation in an iterative manner. Although my cursory impression is that the segmentation/ normalization results look reasonable, I am concerned because I received warning messages for several subjects during the preprocessing (regarding a matrix being close to singular or badly scaled). Based on this warning message, combined with reading correspondence on the SPM list serve [e.g May 23, 2007 (Re: segment error message)], I am concerned that, as John suggested, there were too many Gaussians (4, the default) for modeling the non-brain class and as such there were instabilities in the algorithm. In this regard, I have several questions. 1) Will visual inspection suffice to verify that the images were reasonably registered and segmented? 2) Should I reduce the number of Gaussians used to model non-brain tissue? If so, how many Gaussians should I use? 3) Should I use a mask when segmenting the images? 4) Will problems estimating the distribution for the non-brain tissue class impact the estimation of the distribution of the other classes (e.g. gray matter, white matter, and CSF) either directly or through their impact on bias correction/ registration? 5) If I am concerned that the estimations of my tissue class distributions could be problematic, should I definitely use priors for computing the posterior tissue class probabilities? I appreciate your time and expertise in these matters. Best, Sonya