Bayesian inference of structural brain networks with region-specific Dirichlet parametrisation

In this paper we present an extension to a Bayesian framework for inference of structural brain networks. This framework provides a generative model that explicitely describes how structural brain networks lead to observed streamline distributions. Our extension consists of adding a hyperprior on the latent Dirichlet variables, such that we can capture global and region-specific behaviour within the streamline distributions. We apply these models on both simulated and empirical data. We show that the added flexibility of region-specific parametrisation is not needed for inference of the underlying structure of a brain network, and that the global model with less parameters is still sufficiently flexible to represent data even when generated from a model with region-specific Dirichlet parameters.