Inferring Dynamic Mixtures of Large-Scale Functional Brain Networks Using Deep Learning

COEX 

It is now widely accepted that network activity in the brain changes with time. This talk will focus on methodological advancements for identifying dynamic functional networks in neuroimaging data. Time-varying approaches used to model brain networks often assume a mutual exclusivity over time, e.g. sliding window analyses with clustering methods or the Hidden Markov Model (HMM). Whilst a useful constraint, this assumption may compromise the ability of the approach to describe the data effectively. First, I will review existing methods for identifying dynamic brain networks. Following this, I will introduce a new model for neuroimaging data called DyNeMo (Dynamic Network Modes), which is inspired by recent advances in deep learning. This model surpasses existing methods in two ways: the incorporation of a recurrent neural network capable of modelling long-range structure and the ability to describe the data using a time-varying linear mixture of spatially distributed ‘modes’. We demonstrate DyNeMo’s ability to learn mixtures of networks and to model long-range temporal structure on simulated data. Then applying DyNeMo to real magnetoencephalography data, we show DyNeMo infers plausible functional networks with fast dynamics in resting-state data and networks that reflect actions in task data. We show that DyNeMo provides a complementary description to state-based models. Overall, this is a powerful new approach for studying brain network dynamics in both MEG and fMRI datasets.