Variational inference for medical image segmentation

Variational inference techniques are powerful methods for learning probabilistic models and provide significant advantages over maximum likelihood (ML) or maximum a posteriori (MAP) approaches. Nevertheless they have not yet been fully exploited for image processing applications. In this paper we present a variational Bayes (VB) approach for image segmentation. We aim to show that VB provides a framework for generalising existing segmentation algorithms that rely on an expectation-maximisation formulation, while increasing their robustness and computational stability. We also show how optimal model complexity can be automatically determined in a variational setting, as opposed to ML frameworks which are intrinsically prone to overfitting. Finally, we demonstrate how suitable intensity priors, that can be used in combination with the presented algorithm, can be learned from large imaging data sets by adopting an empirical Bayes approach.