Segmentation

The segmentation in SPM is based on a modified gaussian mixture model (Ashburner & Friston 2000). Bayesian rule is used to assign the probablity for each voxel belonging to each tissue class based on combining the likelihood for belonging to that tissue class and the prior probablity derived from prior probability maps derived from a large number of subjects.

Intensity distribution


intensity-gauss
Segmentation in SPM is obtained from the intensity distribution of the image and prior information for the respective tissue classes. To visualize and analyze image intensity distribution we can plot all values of the image in a histogram. The histogram on the left represents the frequencies of the intensity values of the image. The x-axis shows from left to right increasing image intensities and we can differentiate different curves for each tissue class and background. The right image shows the initial segmentation result without using prior information. If we only use intensity information of the whole image there are several misclassifications outside the cortex. This is because image intensity around the scalp is similiar to that of gray matter as shown in the circles.

Bayes rule to use prior information


Bayes-apriori
We can combine the initial segmentation result with prior information obtained from a large number of correctly segmented images. The prior information outside the cortex shows very low probability for belonging to gray matter. If we combine both probabilities using a Bayesian rule we obtain a joint posterior probability with correct results.