Bias-correction subspace projection method for background noise suppression in OPM-MEG systems

The Magnetoencephalography (MEG) based on optically pumped magnetometers (OPMs) has emerged as a transformative wearable neuroimaging modality, offering significant advantages such as flexible sensor placement and closer proximity to neural sources. However, these very characteristics impose more stringent requirements on the environmental noise suppression. Among spatial filtering approaches, signal space projection (SSP) estimates the noise subspace directly from empty-room recordings and can be implemented without dense spatial sampling, detailed sensor-geometry modeling, or subject-specific forward models. However, as a covariance-based subspace projection method, SSP may yield biased estimates of the noise subspace eigenvectors when the empty-room covariance is contaminated by both environmental interference and sensor noise, leading to incomplete noise removal and potential signal distortion. To address this challenge, this paper proposes a bias-correction subspace projection (BCSP) method. This method constructs a coupling model between sensor noise and the background magnetic field, elucidates the underlying cause of subspace estimation bias in conventional SSP, and estimates bias via least-squares optimization. Combined with an eigenvector correction strategy based on nonlinear soft thresholding, the BCSP method adaptively reduces the bias in the estimated noise subspace eigenvectors, improving background-noise suppression. Both simulation and experimental results consistently demonstrate the proposed method's superior signal-to-noise ratio performance. In auditory evoked field experiments, it robustly preserves the waveform characteristics of the N100 component and generates highly focal cortical activation maps. This method provides a novel approach for noise suppression in OPM-MEG, and holds significant potential to advance the application of this technology in clinical medicine and neuroscience.