Figure 4.

Samples from the approximate posterior distribution of the segmentation p(s|y) when the mode approximation is used in the integral over model parameters (i.e., when the posterior is approximated by Equation 5 as in the baseline segmentation method). The samples can then be obtained by independently sampling the label of each voxel from the categorical distribution in Equation 9. The images correspond to the same subject and slices as in Figure 3; please see its caption for an explanation of the illustration.