XDL-ESI: Electrophysiological Sources Imaging via explainable deep learning framework with validation on simultaneous...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Electroencephalography (EEG) and Magnetoencephalography (MEG) source imaging (ESI) estimates brain activity from scalp recordings.
• The ESI inverse problem is ill-posed, requiring regularization for unique solutions.
• Traditional regularization relies on assumptions about source dynamics, limiting flexibility.

Purpose of the Study:

• To propose a novel Explainable Deep Learning framework (XDL-ESI) for EEG/MEG source imaging.
• To develop a data-driven approach for modeling source solution structure, replacing hand-crafted regularization.
• To enhance robustness and interpretability in ESI.

Main Methods:

• Developed XDL-ESI by integrating iterative optimization with deep learning architecture.
• Unfolded iterative updates into neural network modules.
• Introduced a topological loss leveraging geometric spatial information for improved robustness.

Main Results:

• XDL-ESI provides a data-driven approach, avoiding manual regularization.
• The topological loss enhances source solution robustness by penalizing localization errors.
• The framework demonstrates improved reconstruction efficiency and interpretability.
• Satisfactory performance was achieved on simulated and real clinical data, including simultaneous EEG/iEEG.

Conclusions:

• XDL-ESI offers an efficient, accurate, and interpretable paradigm for solving the ESI inverse problem.
• This deep learning approach advances the field of brain source localization.
• The framework shows significant potential for clinical applications in neuroscience.