Deep learning-based EEG source imaging is robust under varying electrode configurations.

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Conventional electroencephalography source imaging (ESI) requires high-density EEG for accuracy, limiting clinical use.
• Deep learning methods learn spatiotemporal brain activity directly from data, offering potential improvements.
• Low-density EEG is more accessible in clinical settings but traditionally yields less reliable ESI.

Purpose of the Study:

• To evaluate the performance of a novel Deep Learning-based Source Imaging Framework (DeepSIF).
• To assess the impact of varying EEG electrode numbers on DeepSIF's accuracy.
• To compare DeepSIF against conventional ESI methods using both simulated and clinical data.

Main Methods:

• Computer simulations and analysis of clinical data from 27 epilepsy patients were performed.
• EEG source imaging performance was assessed using channel configurations from 16 to 75 electrodes.
• DeepSIF was compared with sLORETA and LCMV conventional methods against ground truth and clinical references.

Main Results:

• DeepSIF demonstrated consistent accuracy in source localization and extent estimation across all tested channel counts and noise levels.
• DeepSIF significantly outperformed conventional methods (sLORETA, LCMV) in accuracy.
• In epilepsy patients, DeepSIF achieved average spatial dispersions of 7.9/9.0 mm (75/16 electrodes), compared to 21.9/28.1 mm for sLORETA and 20.0/28.9 mm for LCMV.

Conclusions:

• The DeepSIF algorithm exhibits robust performance for EEG source imaging even with low-density EEG.
• DeepSIF's effectiveness with fewer electrodes suggests broad clinical applicability.
• This deep learning approach eliminates the need for high-density EEG devices in clinical source imaging.