Calibration-Free Self-Supervised Electromagnetic Interference Elimination for RF Shielding-Free Low-Field MRI Using Repeated-Acquisition Consistency

Abstract

Low-field magnetic resonance imaging (MRI) provides a promising route toward portable and cost-effective imaging, but operation without a radio frequency (RF) shielding room is highly vulnerable to electromagnetic interference (EMI). Existing active EMI elimination methods may require dedicated EMI characterization data or supervised training-data construction using clean MR components and synthetic contaminated-clean pairs. In this work, we propose RACE, a calibration-free self-supervised EMI elimination method for RF shielding-free low-field MRI based on repeated-acquisition consistency. The method exploits the fact that repeated measurements of the same k-space readout have highly consistent underlying MR signals when acquired within a short time interval, whereas their EMI components and random noise may differ. This physical consistency is used to construct a cross-acquisition self-supervised loss, in which the EMI-suppressed signal estimated from one acquisition is cross-validated using the contaminated receive-coil signal and sensing-coil data from the other acquisition. RACE therefore learns directly from EMI-contaminated repeated acquisitions and simultaneously acquired EMI sensing signals, without requiring dedicated EMI characterization data, clean MR labels, or synthetic contaminated-clean training pairs. Reference-based simulation, phantom, and human brain experiments demonstrated that RACE effectively reduced EMI artifacts and improved image quality in RF shielding-free low-field MRI. Compared with External Dynamic InTerference Estimation and Removal (EDITER), RACE achieved higher signal-to-noise ratio (SNR) and noise removal percentage (NRP), while providing performance comparable to Deep learning Direct MR Signal Prediction (Deep-DSP) across the evaluated datasets. These results suggest that repeated-acquisition consistency can serve as an effective self-supervised constraint for practical EMI elimination in low-field MRI.