Feasibility of Implicit Neural Representation Learned Motion Compensation for 3D Stack-of-Spirals Free-Breathing Cardiac Quantitative Susceptibility Mapping.

Purpose Differential blood oxygenation between the right and left heart (ΔSO 2 ) is an indicator of cardiovascular function currently assessed in clinical practice by invasive right heart catheterization. Cardiac MRI can non‐invasively quantify ΔSO 2 with quantitative susceptibility mapping (QSM) using a prospective navigator gated 3D cartesian acquisition. However, this method suffers from long acquisition time and reduced robustness. Here, a free‐breathing cardiac QSM using spiral sampling and deep learning motion compensation is proposed. Methods A retrospective self‐gated stack‐of‐spirals multi‐echo gradient echo sequence is combined with implicit neural representation (INR) learning for image reconstruction. The self‐gating signals measure superior–inferior cardiac and respiratory motion thus allowing k‐space binning. Using a physics‐informed signal model and the spatiotemporal coordinate input, INR infers motion fields as well as motion‐corrected water, fat, and field maps. Then, QSM and ΔSO 2 are accordingly computed. Data were acquired in 10 healthy subjects. For comparison, a free‐breathing prospective navigator ECG‐triggered Cartesian acquisition (NAV) was performed. Results INR reconstructed motion‐corrected water, fat, R2* and field maps were successfully obtained in all subjects. INR‐QSM showed superior image quality ( p = 0.0067) and equivalent ΔSO 2 measurement in the heart ( r = 0.74, p < 0.001; 1.07% ± 3.52% bias/limits of agreement) compared to the reference NAV‐QSM. Conclusion This study demonstrated the feasibility of INR for compensation of cardiac and respiratory motion in free‐breathing 3D cardiac QSM.