 |
 |
After looking at the results of the segmentation, you could score the quality as 0, 1, 2 or 3. This would give a quantitative measure. Alternatively, you could manually segment the images and assess whether the algorithm is giving similar results using something like a kappa statistic. However, I suspect that what you are really after is something a bit more objective, which does not require ground truth based on a subjective opinion from visualising the data (although subjective opinion from visualising scans is all we have for providing the training data used by segmentation algorithms). Although the segmentation algorithms in SPM do not report a measure of model evidence, this would be one approach that could be used for comparing segmentation models. As far as I know, there are not yet any fully Bayesian segmentation approaches that compute a measure of model evidence, so it is a bit tricky to use this measure to compare segmentation models. Also, the measure reflects how well the models encodes the probability density of the data, which I suspect will not really reflect which model most accurately identifies GM. Best regards, -John 2012/2/23 Gustaf Mårtensson <[log in to unmask]>: > Hello all, > > I was wondering if there is any way to obtain a quantitative measure of > the quality of a segmentation performed in SPM? > > Best regards, > > Gustaf Mårtensson > > Neuropediatric Unit > Astrid Lindgren Children's Hospital > [log in to unmask] > Sweden