Bayesian Estimation of Hyperparameters in MRI through the Maximum Evidence Method

Bayesian inference methods are commonly applied to the classification of brain Magnetic Resonance images (MRI). We use the Maximum Evidence (ME) approach to estimate the most probable parameters and hyperparameters for models that take into account discrete classes (DM) and models accounting for the partial volume effect (PVM). An approximate algorithm was developed for model optimization, since the exact image inference calculation is computationally expensive. The method was validated using simulated images and a digital phantom. We show that the Evidence is a very useful figure for error prediction, which is to be maximized respect to the hyperparameters. Additionally, it provides a tool to determine the most probable model given measured data.