Michael Hütel1,2, Andrew Melbourne1, Jonathan Rohrer2, and Sebastien Ourselin1,2
1Translational Imaging Group, University College London, London, United Kingdom, 2Dementia Research Centre, University College London, London, United Kingdom
Independent Component Analysis (ICA) has been proven to produce compact representations of recurrent patterns in BOLD-fMRI imaging data. Most ICA implementations used in BOLD-fMRI, however, optimize for spatial sparse decompositions rather than independent decompositions. We describe a neural-network ICA framework that optimizes directly for sparsity and also allows for overcomplete basis representation.
