Reliable Off-Resonance Correction in High-Field Cardiac MRI Using Autonomous Cardiac B0 Segmentation with Dual-Modality Deep Neural Networks

B0 field inhomogeneity is a long-lasting issue for Cardiac MRI (CMR) in high-field (3T and above) scanners. The inhomogeneous B0 fields can lead to corrupted image quality, prolonged scan time, and false diagnosis. B0 shimming is the most straightforward way to improve the B0 homogeneity. However, today's standard cardiac shimming protocol requires manual selection of a shim volume, which often falsely includes regions with large B0 deviation (e.g., liver, fat, and chest wall). The flawed shim field compromises the reliability of high-field CMR protocols, which significantly reduces the scan efficiency and hinders its wider clinical adoption. This study aims to develop a dual-channel deep learning model that can reliably contour the cardiac region for B0 shim without human interaction and under variable imaging protocols. By utilizing both the magnitude and phase information, the model achieved a high segmentation accuracy in the B0 field maps compared to the conventional single-channel methods (Dice score: 2D-mag = 0.866, 3D-mag = 0.907, and 3D-mag-phase = 0.938, all p < 0.05). Furthermore, it shows better generalizability against the common variations in MRI imaging parameters and enables significantly improved B0 shim compared to the standard method (SD(B0Shim): Proposed = 15 ± 11% vs. Standard = 6 ± 12%, p < 0.05). The proposed autonomous model can boost the reliability of cardiac shimming at 3T and serve as the foundation for more reliable and efficient high-field CMR imaging in clinical routines.