A Deep Learning Framework for Multi-Source EEG Localization.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Electroencephalography (EEG) offers high temporal resolution but limited spatial accuracy for multiple neural sources.
• Classical inverse methods often fail to localize closely spaced or weak neural generators due to "single-source bias".

Purpose of the Study:

• To develop a deep learning framework for robust multi-source localization from short EEG segments.
• To overcome limitations of traditional EEG source imaging techniques.

Main Methods:

• A convolutional neural network (ConvNET) was trained on realistic EEG simulations.
• A distinct forward model was used during training to prevent "inverse crime" and ensure generalization.
• The ConvNET was benchmarked against nine established inverse solvers.

Main Results:

• The deep learning approach consistently outperformed traditional solvers in resolving closely spaced sources.
• Accuracy was maintained or improved for single-source localization compared to existing methods.
• The framework demonstrated superior performance across various synthetic test scenarios.

Conclusions:

• Deep learning offers a more reliable method for EEG source localization, overcoming biases inherent in classical approaches.
• This advancement has significant potential for applications in presurgical planning, brain-computer interfaces, and neurofeedback.