| Crynodeb: | <p>Neuroimaging data is often high-dimensional and difficult to interpret. Methods have been developed which can be applied to datasets to make them more malleable and comprehensible to researchers. This process is critical for improving our understanding of the nature of the brain. The developments made through neuroscientific research have the potential to transform the lives of many people, and making progress in the development of new methods is an important part of the discovery pipeline.</p>
<p>In this thesis, a novel recurrent neural network model, (DyNeMo), is introduced to improve on the neuroimaging analysis performance (in particular (MEG) data) compared to the current gold standard analysis method, the hidden Markov model. (HMMs) depend on the Markovian property and mutual exclusivity of activating states, ignoring long-range structure within data. DyNeMo removes these limitations to allow for a potentially deeper analysis.</p>
<p>The validity of DyNeMo was assessed with a series of simulations to test the performance of DyNeMo in comparison to the HMM. As expected, the HMM and DyNeMo both performed well for HMM and HSMM simulated data. For a more complex simulation of a linear mixture of several modes, DyNeMo outperformed the HMM. The HMM was unable to detect co-activation of states and learn long-range dependencies of data, whereas DyNeMo could.</p>
<p>DyNeMo was applied to real-world MEG data from 51 subjects. DyNeMo revealed a new interpretation of the data. The modes of DyNeMo were found to have modelled the structure of a task performed by the participants which was hidden to the model during training. It also successfully captured the spectral content of the data.</p>
<p>In an alternative regime, DyNeMo was applied to each channel of the dataset in isolation. By selecting functions which optimise some aspect of the data, low dimensional modes can be created by combining these single-channel modes. It was found that different interpretations of the data could be produced from the same models. This approach also has the advantage of a reduced memory requirement during training, making it more accessible to researchers.</p>
<p>DyNeMo outperforms its predecessors. Its adoption could improve uptake of MEG in research and clinical settings and ultimately improve patient outcomes and lives in numerous neurological conditions.</p>
|
|---|