Subspace Implicit Neural Representations for Real-Time Cardiac MR Imaging

Motivation: Existing cardiac cine MRI methods achieve limited temporal resolution due to retrospective gating, which limits accurate capture of continuous cardiac dynamics. Goal(s): To develop a reconstruction framework for real-time cardiac cine MRI using subspace implicit neural representations, aiming to improve both spatial and temporal resolution. Approach: We use two MLPs to learn the spatial and temporal subspace bases, leveraging the low-rank properties of cardiac cine MRI. Networks are initialized with low-resolution GRASP reconstruction and fine-tuned using spoke-specific losses to recover details. Results: Our real-time approach achieves superior quality compared to NUFFT and GRASP reconstruction, which were evaluated across several temporal resolutions. Impact: This method provides a novel reconstruction approach for real-time cardiac MRI with continuous radial acquisition, and will potentially reducing scan times and improving diagnostic capabilities, especially for imaging arrhythmias and characterizing beat-to-beat dynamics compared to conventional approaches.