A Self-Attention-Based Model for Automatic Magnetoencephalography Signal Reconstruction

Magnetoencephalography (MEG) is a non-invasive neuroimaging technique with high spatiotemporal resolution, widely utilized in neuroscience and clinical research. However, noise interference and physiological activities often result in bad channels and segments of MEG data, adversely affecting subsequent analysis. Most common methods directly reject or interpolate to repair these channels and segments, which lead to data loss or significant signal distortion. Here, we propose a self-attention-based MEG reconstruction model capable of reconstructing signals in corrupted channels and segments that closely match the real data in both time and frequency domains, while enhancing the data quality without the need for manual feature design. Even in the presence of numerous bad channels and segments, our model consistently achieves high-quality signal reconstruction. Overall, we introduce a novel deep learning strategy for the automatic repair of MEG signals, demonstrating the tremendous potential of artificial intelligence in improving neuroimaging data analysis.