A physics-driven deep unrolling framework with multiscale attention for multichannel Magnetocardiography denoising

Magnetocardiography (MCG) measurements are often contaminated by strong non-stationary interference, which degrades the fidelity and repeatability of the acquired signals. To address this problem, we propose Magnetocardiography-Alternating Direction Method of Multipliers-Net (MCG-ADMM-Net), a physics-driven deep unrolling framework with multiscale attention for multichannel Spin-Exchange Relaxation-Free Magnetocardiography (SERF-MCG) denoising. The network unrolls the ADMM optimization procedure into a cascade of iterative stages and combines a physically constrained reconstruction branch, a multiscale attention-based deep prior, an adaptive parameter determination module, and a gated residual fusion strategy. A composite loss function is further introduced to enforce reconstruction fidelity, physical consistency, energy-aware recovery, and morphological preservation. Experiments on a multichannel MCG dataset showed that the proposed method consistently outperforms conventional filtering approaches, data-driven deep learning models, and existing deep unrolling methods. Under standard noise conditions, MCG-ADMM-Net achieved a Signal-to-Noise Ratio improvement ( ) of 20.15 dB, a Percentage Root Mean Square Difference (PRD) of 9.87%, and a Correlation Coefficient (CC) of 0.99. Performance remained stable across a wide range of noise intensities, indicating strong robustness under severe interference. In addition, validation on a public Electrocardiography (ECG) dataset constructed from QT Database (QTDB) and MIT-BIH Noise Stress Test Database (NSTDB) confirmed good cross-dataset generalization. By effectively bridging the gap between physical model-based optimization and data-driven deep learning, these results suggest that MCG-ADMM-Net is an effective and interpretable solution for high-fidelity multichannel MCG denoising.