LG-BiTCN: High-Fidelity Denoising for MCG in Strong Noise

Abstract Objective. This study investigates the denoising of low-cost magnetocardiography (MCG) signals recorded under strong noise conditions. Approach. We propose LG-BiTCN (Least-Squares Generative Adversarial Network with Gated Bidirectional Temporal Convolutional Network), which combines long-range temporal feature extraction with adversarial training for signal denoising. Using clean MCG signals from the Kiel Cardio Database, we design a composite noise model consisting of baseline drift, 1/f (pink) noise, and white Gaussian noise. Main results. In all composite noise conditions, LG-BiTCN achieves the best denoising performance. At −20 dB input signal-to-noise ratio (SNR) with baseline drift + 1/f noise + white noise, LG-BiTCN improves SNR by 24. 21 dB, outperforming traditional algorithms by more than 8. 84 dB. Additionally, LG-BiTCN demonstrates superior waveform fidelity, as reflected by higher SSIM and lower MAEₐₑₒ compared to baseline methods. We find that at very low SNR, larger receptive field designs are more beneficial for improving denoising performance, while at higher SNR, smaller receptive fields better preserve signal details. Significance. These results demonstrate that LG-BiTCN can effectively enhance MCG signal denoising under high-noise conditions, providing valuable insights for methods in unshielded MCG denoising tasks.