Generative MR Multitasking: Bridging real-time and gated imaging with complex-valued cardiac encoding and neural subspace representations

Motivation: To improve representation of cardiac and respiratory motion for multidimensional imaging and bridge the divide between gating and real-time imaging. Goal(s): To develop continuous representations of respiratory/cardiac motion, and extend the standard cardiac phase paradigm to a complex-harmonic inspired cardiac phase (timing) and "amplitude" (beat-to-beat variability) paradigm. Approach: We propose a generative-subspace MR Multitasking framework using a conditional variational autoencoder (CVAE) for an implicit neural representation of continuous-time multidimensional temporal subspaces. Results: The CVAE subspace produced higher SNR, continuous motion, and eliminated eddy current artifacts. The complex cardiac representation simultaneous allowed both "real-time" beat-by-beat imaging and average-cycle "gated" imaging. Impact: This framework bridges traditional gating and real-time imaging approaches, and provides a flexible, scalable framework for multidimensional imaging. The complex-harmonic cardiac representation preserves both cardiac timing (phase) and beat-to-beat variability (amplitude), which may especially have future impact for arrhythmia patients.