A Bayesian Deep CNN Framework for Reconstructing k-t-Undersampled Resting-fMRI

Abstract

Undersampled reconstruction in resting functional magnetic resonance imaging (R-fMRI) holds the potential to enable higher spatial resolution in brain R-fMRI without increasing scan duration. We propose a novel approach to reconstruct kt undersampled R-fMRI relying on a deep convolutional neural network (CNN) framework. The architecture of our CNN framework comprises a novel scheme for R-fMRI reconstruction that jointly learns two multilayer CNN components for (i) explicitly filling in missing k-space data, using acquired data in frequency-temporal neighborhoods, and (ii) image quality enhancement in the spatiotemporal domain. The architecture sandwiches the Fourier transformation from the frequency domain to the spatial domain between the two aforementioned CNN components, during, both, CNN learning and inference. We propose four methods within our framework, including a Bayesian CNN that produces uncertainty maps indicating the per-voxel (and per-timepoint) confidence in the blood oxygenation level dependent (BOLD) time-series reconstruction. Results on brain R-fMRI show that our CNN framework improves over the state of the art, quantitatively and qualitatively, in terms of the connectivity maps for three cerebral functional networks.