This dissertation presents a new approach to stochastic modeling of magnetic resonance images. A rigorous and comprehensive model of MR image formation is derived based on the physics of magnetic resonance image formation, and the model is validated using real MRI data. A general theoretical result about the existence of spatial autoregressive processes is presented. We attempted to obtain an accurate tissue classification map for a set of real MR images based on two doctors' hand tracings of the boundaries of the regions of the different tissues in the images. In accordance with the prior knowledge of the tissue map, we designed a prior model for the distribution of the underlying tissue map. The technique is applied to tissue classification on a set of real MR images by the use of the Bayesian formulation and ICM algorithm.