GETNet: a gate-enhanced temporal network for electrophysiological source imaging

Accurately characterizing the spatiotemporal distribution of brain activity from scalp electroencephalography (EEG) is essential for advancing our understanding of brain function. However, traditional electrophysiological source imaging (ESI) methods rely heavily on hand-crafted priors and are sensitive to noise and parameter settings, which limits their robustness and generalizability. In this study, we present GETNet (Gate-Enhanced Temporal Network), a data-driven framework that learns the mapping from scalp EEG signals to cortical source activity through cascaded spatiotemporal modules. The framework integrates an extended gated residual network (EGRN) for channel expansion and noise suppression, a temporal module combining multi-head self-attention (MHA) and bidirectional LSTM (BiLSTM) for capturing long-range and local dependencies, and a gated linear unit (GLU) for generating physiologically plausible sparsity at the output. To comprehensively evaluate the performance of the method, we conducted extensive experiments across diverse simulation scenarios, including varying activation extents, signal-to-noise ratios, and forward-model mismatches. The results indicate that GETNet consistently achieved superior accuracy and stability in reconstructions compared to both traditional approaches and recent neural network frameworks. Furthermore, we also validated the effectiveness of GETNet on two real-world datasets, indicating that it has broad applicability for neuroscience research.