Electrophysiological brain imaging based on simulation-driven deep learning in the context of epilepsy

Area of Science:

• Neuroscience
• Medical Imaging
• Computational Biology

Background:

• Electrophysiological Source Imaging (ESI) is crucial for localizing brain activity in epilepsy but is an ill-posed inverse problem.
• Existing ESI methods often rely on physiological constraints to ensure unique solutions.
• Accurate localization of epileptic sources from high-resolution ElectroEncephaloGraphy (Hr-EEG) is challenging due to signal complexity and noise.

Purpose of the Study:

• To propose an efficient, patient-specific Electrophysiological Source Imaging (ESI) approach leveraging simulation-driven deep learning.
• To accurately identify the location, spatial extent, and electrical activity of distributed brain sources in epilepsy.
• To enhance the robustness and performance of ESI methods for clinical applications in presurgical evaluation.

Main Methods:

• A novel ESI approach based on simulation-driven deep learning using realistic, patient-specific simulated epileptic High-resolution 256-channels scalp EEG (Hr-EEG) signals.
• Utilized a computational neural mass model for temporal dynamics and a patient-specific head model with the boundary element method for the forward problem.
• Employed a Temporal Convolutional Network (TCN) for spatial patterns and Long Short-Term Memory (LSTM) for temporal dependencies, incorporating a multi-scale strategy.

Main Results:

• The proposed method achieved a Dipole Localization Error (DLE) of 1.39 and Normalized Hamming Distance (NHD) of 0.28 for single-source scenarios (SNR 10 dB).
• For two uncorrelated sources (SNR 10 dB), DLE was 1.50 and NHD was 0.28; for two correlated sources, DLE was 3.74 and NHD was 0.43.
• Demonstrated superior performance compared to two deep learning and four classical ESI methods on both simulated and real interictal Hr-EEG data, showing robustness to noise.

Conclusions:

• The simulation-driven deep learning ESI approach is an efficient and accurate tool for localizing epileptic brain areas from Hr-EEG.
• The proposed method outperforms existing techniques, offering a promising alternative for reconstructing electrical activity in epilepsy.
• Its robustness to noise makes it a valuable tool for presurgical evaluation and understanding epilepsy.