A Novel End-to-End Framework to Image Cortical Networks from EEG

Deciphering the intricate causal relationships among cortical sources in the brain's complex systems through a non-invasive and low-cost technique remains a formidable challenge. Electroencephalography (EEG), with its high temporal resolution, offers a promising avenue to unravel the underlying neural mechanisms. To this end, a novel end-to-end model is proposed to capture cortical networks from EEG data based on a temporal convolutional graph neural network (GNN), called TCGNN-BrainNet. The core components of this model are the inverse operator neural network (IONN) and the connectivity estimation neural network (CENN). The IONN transforms the EEG signal into an abstract representation of deep cortical activity, while the CENN constructs and leverages graph encoding for interaction inference. Through extensive training, TCGNN-BrainNet acquires the capability to infer cortical networks directly from EEG data, optimizing accuracy and mitigating error propagation common in conventional methods. Notably, the training strategy concurrently weaves the understanding of physiological cortical features with the implicit relational understanding of supervisory data, through the defined loss function. This enables the neural networks to organically uncover and seize subtle interaction patterns amidst the signals. Extensive simulations were performed to substantiate the efficacy of our model, evidencing the robustness and consistent reliability of TCGNN-BrainNet in accurately constructing cortical networks. These promising outcomes demonstrate the significant potential of TCGNN-BrainNet in practical brain network analyses, paving the way for new frontiers in neuroscientific exploration and advancements in the field.