Generative MR Multitasking with complex-harmonic cardiac encoding: Bridging the gap between gated imaging and real-time imaging

TL;DR

This paper introduces a generative neural network framework that unifies gated and real-time cardiac MRI, enabling flexible motion representation, artifact suppression, and improved quantitative mapping in a single free-breathing, non-ECG-gated acquisition.

Contribution

It presents a novel Generative Multitasking approach using a CVAE with complex harmonic cardiac coordinates to bridge the gap between gated and real-time cardiac MRI.

Findings

Reconstructed cardiac phase-resolved and time-resolved series.

Reduced intraseptal T1 and T2 coefficient of variation, indicating higher SNR.

Suppressed trajectory-dependent artifacts in reconstructions.

Abstract

Purpose: To develop a unified image reconstruction framework that bridges real-time and gated cardiac MRI, including quantitative MRI. Methods: We introduce Generative Multitasking, which learns an implicit neural temporal basis from sequence timings and an interpretable latent space for cardiac and respiratory motion. Cardiac motion is modeled as a complex harmonic, with phase encoding timing and a latent amplitude capturing beat-to-beat functional variability, linking cardiac phase-resolved ("gated-like") and time-resolved ("real-time-like") views. We implemented the framework using a conditional variational autoencoder (CVAE) and evaluated it for free-breathing, non-ECG-gated radial GRE in three settings: steady-state cine imaging, multicontrast T2prep/IR imaging, and dual-flip-angle T1/T2 mapping, compared with conventional Multitasking. Results: Generative Multitasking provided flexible cardiac motion representation, enabling reconstruction of archetypal cardiac phase-resolved cines (like gating) as well as time-resolved series that reveal beat-to-beat variability (like real-time imaging). Conditioning on the previous k-space angle and modifying this term at inference removed eddy-current artifacts without globally smoothing high temporal frequencies. For quantitative mapping, Generative Multitasking reduced intraseptal T1 and T2 coefficients of variation compared with conventional Multitasking (T1: 0.13 vs. 0.31; T2: 0.12 vs. 0.32; p<0.001), indicating higher SNR. Conclusion: Generative Multitasking uses a CVAE with complex harmonic cardiac coordinates to unify gated and real-time CMR within a single free-breathing, non-ECG-gated acquisition. It allows flexible cardiac motion representation, suppresses trajectory-dependent artifacts, and improves T1 and T2 mapping, suggesting a path toward cine, multicontrast, and quantitative imaging without separate gated and real-time scans.