Use of tissue priors (experimental!)

SPM5 uses priors with a Bayes rule to combine the likelihood for belonging to a tissue class and the prior probability derived from prior probability maps. However, the use of priors might cause problems for brains deviating from a healthy, adult control population. The ICBM priors in SPM5 are based on brains of subjects with an mean age of 25 years. Thus, brains of children or or elderly subjects will deviate from the tissue probabilities of the ICBM priors and will bias the segmentation. This options avoids the dependency on tissue priors and usually increases classification accuracy. Segmented images show clearer delineation particularly in the basal ganglia and in the sulci and this is the prefered option for most images. If no priors are used the segmented images are indicated by an “p” instead of “c”.
Keep in mind that even if you are using this option tissue priors are always used to register your images to MNI space. Only the bayesian rule to combine priors with tissue type probabilities will be omitted. Thus, for some cases the use of customized templates might utilize the spatial registration (e.g. infants).
However, there are some cases (e.g. large atrophy), where this option fails and gray matter distribution might be underestimated. You will notice these cases by a very thin gray matter ribbon and overestimated CSF. If this happens the default SPM5 tissue priors (or own priors) should be used.

Example of an infant brain


NoPriors
The upper row shows the T1 image (a) of an infant brain (courtesy of Scott Holland) and the SPM5 segmentation using tissue priors (b-d). Fig. 2e shows the difference between the GM segmentation with and without tissue priors. Blue colors indicate that GM probability of the new approach is larger than the default SPM5 method and red colors point to larger GM probabilities of the default SPM5 segmentation. The new segmentation without use of tissue priors is displayed in f-h.

Comparison with default segmentation


NoPriors-Kappa
To validate the effect of not using the tissue priors we use a ground truth image from the brainweb database with varying noise levels of 1..9%. We have calculated the kappa coefficent as measure of agreement. A kappa coefficient of 1 means that there is perfect overlap between the ground truth reference and the segmented image. The proposed approach without using priors, but considering MRF achieve best results over a wide range of noise levels. Only for data with more than 7% noise the MRF approach outperforms our method.