Generalizing Diffusion Tensor Model Using Probabilistic Inference in Markov Random Fields

We give a proof of concept for efficiently modeling tensor configuration distributions with Markov random fields (MRFs) and inferring the most likely tensor configurations with maximum a posteriori (MAP) estimations. We demonstrate the plausibility of our method by resolving fiber crossings in a synthetic dataset, experimenting with three different MAP estimation methods on a grid MRF model. The power of the MAP-MRF framework comes from its mathematical convenience in modeling prior distributions and the fact that it yields a global optimization driven by local neighborhood interactions.