Bayesian reconstruction of R-fMRI from K-T undersampled data using a robust, subject-invariant, spatially-regularized dictionary prior

Faster resting-state functional magnetic resonance imaging (R-fMRI) can improve spatiotemporal resolution and functional sensitivity. To speedup scans, current methods rely on complex pulse-sequence design or straightforward undersampling along with (weak) priors on the signal. We propose a Bayesian graphical R-fMRI reconstruction framework relying on learning data-adaptive prior models through dictionaries that we design to be robust to large physiological fluctuations typical in R-fMRI signals. Our dictionary adapts to multiple subjects through an optimal similarity transform. Our reconstructions on simulated and real-world R-fMRI give more accurate functional networks and better spatial resolution than the state of the art.