Unsupervised reconstruction of highly undersampled 3D cones cardiac image navigators using a dual-branch joint training framework

Motivation: Image-based navigators (iNAVs) are pivotal for motion correction in cardiac MRI. High undersampling rates and absence of ground truth poses challenges to 3D iNAVs reconstruction. Goal(s): To develop an unsupervised approach for 3D iNAV reconstruction from highly undersampled non-Cartesian k-space data. Approach: We employed a dual-branch joint training framework to implement cross self-supervised learning between the two reconstruction branches. Further expansion learning was performed on augmented reconstructed iNAVs. A feature-level unrolled model was employed for reconstruction. Results: For 10 subjects, our approach reconstructed 3D iNAVs from 12x undersampled data acquired in each heartbeat, demonstrating superior reconstruction quality compared to self-supervised and supervised methods. Impact: Our approach provides a way to reconstruct highly undersampled 3D cardiac images with sufficient quality for retrospective motion correction, suggesting the utility of our approach in various scenarios where fully sampled data is unavailable.