M/EEG source localization for both subcortical and cortical sources using a convolutional neural network with a...

Area of Science:

• Neuroscience
• Medical Imaging
• Computational Biology

Background:

• Electroencephalography (EEG) and magnetoencephalography (MEG) offer noninvasive brain signal measurement but have limited spatial resolution.
• Electrophysiological source imaging (ESI) aims to noninvasively identify the neuronal sources of M/EEG signals.
• Accurate localization of subcortical brain sources using M/EEG remains a significant challenge in neuroscience and clinical practice.

Purpose of the Study:

• To develop and validate a data-driven, deep learning-based ESI approach for precise localization of both cortical and subcortical brain sources.
• To overcome limitations of traditional ESI methods that rely on explicit regularization priors.
• To establish the feasibility and accuracy of deep learning for subcortical source localization in M/EEG analysis.

Main Methods:

• A four-layered convolutional neural network (4LCNN) was designed for ESI without requiring predefined regularization priors.
• A realistic head conductivity model was created using individual MRI data with advanced segmentation of ten distinct head tissues.
• The 4LCNN model was trained using simulated M/EEG data generated from the realistic head model.

Main Results:

• The proposed 4LCNN method demonstrated high accuracy in localizing electrophysiological sources, notably outperforming existing methods in subcortical regions.
• Validation was confirmed through M/EEG simulations, analysis of evoked responses, and comparison with invasive recordings.
• This deep learning approach represents the first successful application of ESI targeting subcortical brain areas.

Conclusions:

• The developed deep learning-based ESI method, particularly the 4LCNN, offers a significant advancement in accurately localizing subcortical and cortical brain sources from M/EEG data.
• This technique holds substantial potential for enhancing clinical diagnosis, understanding neuronal disease pathophysiology, and advancing basic brain function research.
• The study validates the feasibility and superior performance of deep learning for challenging subcortical source localization tasks.