Before we start defining the MRF model, let us summarize the notation that we will be using in the remainder of the chapter. In our notation, we will primarily follow Besag's treatment [10]. Let I denote a 2-D rectangular lattice (i.e., our image) of Mx χ Μ pixels. Note that a 3-D lattice (or cubic lattice) will have Mx χ Μ χ Mz voxels whose center locations
will lie at the lattice points. We assume that pixel or voxel г will take on values (i.e., labels) x. from the set L = [1,···Κ], where К denotes the total number of regions available in the image because labels represent regions in image I. Let us assume that x* is the true labeling of the pixels or voxels in the image. That is, our true image x* is a realization of the random field X. The sample space Ω denotes the space of all possible realizations of our random field X. That is, Ω = {1,...,Κ} x у for 2-D images, and Q={l,...,iqM*XM*XM4or3-D images. Recall the multiplication principle in probability, mentioned at the beginning of the chapter, used to compute the size of the state space. Let us suppose we have three regions (i.e., К = 3). In this case, each pixel has three possible labels. If we suppose that we are working on a series of 512 χ 512 MR images with three regions, the size of the sample space in this case would be 3512x512 = 9.8936 χ 10246. This number illustrates clearly the magnitude of the problem that we have to deal with.