Regularisation Using Spatiotemporal Independence and Predictability

Independent component analysis (ICA) of an image sequence extracts a set of statistically independent images, and deenes a corresponding set of unconstrained dual time courses. However, the extra degrees of freedom implicit in these time courses can lead to physically improbable solutions. Accordingly, we introduce two methods for regularising ICA: smoothed independent component analysis (smICA), and spatiotemporal ICA (stICA). smICA is based on the observation that temporal signals tend to vary smoothly, and stICA is based on the observation that images and time courses tend to be statistically independent. A key feature of these methods is that they are based on generic physical properties , and may therefore be widely applicable. The methods are demonstrated on synthetic fMRI images.