| Summary: | <p>The human brain is a highly dynamical system. One way of imaging neural activity is through electrophysiological techniques such as magnetoencephalography and electroencephalography (M/EEG). However, for these data to be truly useful, the ill-posed inverse problem must first be tackled.</p>
<p>A fundamental step in solving this problem involves the introduction of assumptions about the data (and noise) covariance. To date, the majority of source localisation algorithms employ static approximations to the data covariance matrix.</p>
<p>Here we make the argument that researchers should consider a Temporally Adaptive SourcE Reconstruction (TASER) approach to solving the inverse problem. We achieve this by introducing a novel machine learning unsupervised recurrent neural network (RNN) based algorithm. Our model does not make restrictive assumptions about the temporal processes generating the data (i.e. it is not Markovian), nor does it assume that data originate from a set of mutually exclusive states. Instead, we model the data as a continuous mixture of temporally evolving modes or elements.</p>
<p>This algorithm, which we call VRAD, lets us add a temporal extension onto preexisting M/EEG source reconstruction approaches. In this thesis we develop and explore the concepts of temporally adaptive spatial filters (beamformers) and generative model-based reconstruction procedures (e.g. multiple sparse priors). This is motivated first by simulations which show an improvement in source reconstruction accuracy when a time-varying beamformer is used, and the ground truth source space data has a dynamic covariance profile. We then validate the algorithm on a previously published MEG head cast dataset Barratt et al. (2018). Our results show an improvement upon existing static techniques, and we are able to spatially distinguish activation on the order of a few millimetres between pinky and index finger movement. Despite never being exposed to the underlying task-structure, we show that VRAD is capable of learning physiologically plausible latent descriptions of the data. We further assess the philosophy of TASER on a more generic multi-modal multi-subject openly available M/EEG dataset published by Wakeman and Henson.
We show that using multiple source reconstruction weights is not detrimental to the reconstruction of neural waveforms, and that dynamic techniques have benefits of static reconstruction algorithms.</p>
|
|---|