Abstracts

Rationale: Noninvasive electromagnetic recording techniques such as magnetoencephalography (MEG) are commonly used in the routine presurgical evaluation of drug resistant epilepsy (DRE) patients. Underlying brain electrical activity from MEG recordings can be estimated using the electrophysiological source imaging (ESI) technique. ESI has been widely used to help identify epileptogenic tissue by analyzing the interictal spikes or the ictal signals. Conventional ESI methods, however, require explicitly defining priors and regularization terms. This introduces complications such as tuning hyperparameters, making it challenging for the conventional ESI methods to provide the accuracy, efficiency, and objectivity needed for guiding treatment planning in DRE patients. On the other hand, the deep learning-based ESI method can learn implicit prior information regarding the brain sources from a large amount of training data.

Methods: We have developed a novel Personalized Deep learning-based Source Imaging Framework (P-DeepSIF). Realistic synthetic brain activities and the corresponding MEG signal are generated using brain dynamical models and patient-specific head models to train a neural network for the ESI problem. The network contains three blocks of two fully connected layers with a skip connection, and Exponential Linear Unit (ELU) activation function, and three Long Short Term Memory (LSTM) layers with hyperbolic tangent (TanH) activation units. P-DeepSIF was trained with 600,000 samples and the test dataset includes a total of 20,000 examples. In addition to simulated test data, the trained model was evaluated on the interictal spikes from 5 focal epilepsy patients who underwent intracranial EEG implantation and became seizure-free after surgery with at least 1 year of follow-up. Imaging results were compared to clinically defined seizure onset zone (SOZ) from intracranial EEG recordings. Standardized low-resolution brain electromagnetic tomography (sLORETA) was used as the benchmark to compare with the P-DeepSIF’s results.

Results: P-DeepSIF has a significantly smaller localization error compared to sLORETA when evaluated on the simulation dataset (Figure 1A, two tailed paired T-test, p< 0.01). DeepSIF can also estimate the size of the underlying sources with high accuracy. On the clinical dataset, P-DeepSIF (12.51 < ![if !msEquation] > < ![endif] > 2.64 mm) shows superior performance for localizing interictal spike signals compared to sLORETA (21.80 < ![if !msEquation] > < ![endif] > 4.59 mm - Figure 1B).

Conclusions: We have developed a novel data-driven dynamic functional source imaging framework based on deep learning. P-DeepSIF demonstrates superior performance for localizing and imaging epileptogenic tissue from interictal spikes, suggesting the potential for aiding clinical diagnosis and treatment of epilepsy.

Funding: This work was supported in part by NIH NS096761, EB021027, and by a gift from the Pittsburgh Health Data Alliance. R.S. was supported in part by a Fellowship from the Center for Machine Learning and Health.