Electromagnetic Source Imaging via a Data-Synthesis-Based Convolutional Encoder-Decoder Network.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Electromagnetic source imaging (ESI) is crucial for understanding brain activity but faces challenges due to its ill-posed inverse nature.
• Traditional ESI methods rely on predefined priors that may not accurately represent real brain sources, limiting their applicability.

Purpose of the Study:

• To introduce a novel data-synthesized spatiotemporally convolutional encoder-decoder network (DST-CedNet) for robust electromagnetic source imaging.
• To overcome the limitations of conventional ESI techniques by integrating advanced machine learning approaches.

Main Methods:

• The DST-CedNet framework reframes ESI as a machine learning task, utilizing discriminative learning and latent-space representations.
• A unique data synthesis strategy is employed, incorporating prior knowledge of dynamic brain activities to generate extensive training datasets for the convolutional encoder-decoder network (CedNet).

Main Results:

• The DST-CedNet demonstrated superior performance compared to state-of-the-art ESI methods in numerical simulations.
• Analysis of real magnetoencephalography (MEG) and epilepsy electroencephalography (EEG) datasets confirmed the robust estimation of source signals across diverse configurations.

Conclusions:

• The proposed DST-CedNet offers a powerful and robust alternative for electromagnetic source imaging.
• This machine learning-based approach effectively learns brain activity mapping from E/MEG signals, outperforming traditional methods.