A spatially variant mixture model for diffusion weighted MRI: application to image denoising

High angular resolution diffusion imaging is an increasingly important image modality. The nature of the diffusion data makes mixtures of probability distributions particularly suitable for modeling its signals. In this paper, we introduce Bayesian finite mixture models for studying the diffusion field. We apply a spatially variant mixture model to study prior distributions on the model parameters of the field. The mean vectors and covariance matrices are independent of lattice locations, but the mixture weights are allowed to differ from one location to another. Spatial smoothness is achieved by placing a Markov random field prior on top of the mixture weights. The output is a general model that can be used in different HARDI applications, such as fiber tracking and image denoising. The latter is illustrated in this study, with promising results on a real dataset.