A vector Bayesian beamforming framework with noise covariance learning for robust MEG source imaging.

Magnetoencephalography (MEG) provides unparalleled spatiotemporal resolution for non-invasive brain functional imaging. This study aims to improve the robustness and accuracy of voxel-based vector beamformers for MEG source imaging by addressing covariance estimation instability under low-SNR, limited-sample, correlated-source, and non-white noise conditions. In this work, we propose a vector Bayesian beamforming framework with noise covariance learning to address these challenges. The proposed method formulates MEG source imaging as a hierarchical Bayesian inference problem, jointly modeling vector sources and full positive-definite noise covariance structures. By leveraging data-driven covariance learning and sparsity-promoting priors, the framework enables adaptive refinement of the data covariance matrix, thereby enhancing spatial focusing and temporal reconstruction fidelity. Notably, the proposed approach provides a unified interpretation of covariance regularization in beamforming and improves robustness against correlated sources and noise contamination. We evaluate the method on both simulated correlated sources datasets across a wide range of SNR conditions and real 64-channel optically pumped magnetometer (OPM)-MEG recordings under multiple stimulus paradigms. Experimental results demonstrate that the proposed approach achieves an average improvement of 18.03% in AUC compared to conventional beamformers, while maintaining millimeter-level localization accuracy. These results highlight its superior performance in both spatial localization and dynamic signal reconstruction, suggesting strong potential for neuroscience research and clinical applications. The implementation is publicly available at: https://github.com/gao815/VBNLBF.