Spatiotemporal Rhythmic Seizure Sources Can be Imaged by means of Biophysically Constrained Deep Neural Networks.

Noninvasive dynamic brain imaging of neural oscillations provides valuable insights into both physiological and pathological brain states. Yet, challenges remain due to the ill-posed nature of the problem and high complexity of the solution space, which can be alleviated by advanced computational models. Here, we investigated the capability of a novel deep learning-based source imaging framework (DeepSIF) for imaging ictal activities from high-density electroencephalogram (EEG) recordings in drug-resistant focal epilepsy patients. The neural mass model of ictal oscillations was adopted to generate synthetic training data with spatio-temporal-spectra features similar to ictal dynamics. We rigorously validated the trained DeepSIF model using computer simulations and in a cohort of 33 drug-resistant focal epilepsy patients. The DeepSIF ictal source imaging was compared with interictal source imaging and three conventional imaging methods as benchmark comparisons. Our findings show that the trained DeepSIF model outperforms other methods in estimating the spatial and temporal information of ictal sources. It achieves a high spatial specificity of 96% and a low spatial dispersion of 3.80 ± 5.74 mm when compared to the resection region. The noninvasive source imaging results also demonstrate good coverage of seizure-onset-zone (SOZ), with an average distance of 10.89 ± 10.14 mm (from the SOZ to the reconstruction). These promising results suggest that DeepSIF has significant potential for advancing noninvasive imaging of ictal activities in patients with focal epilepsy. By providing valuable insights into the spatiotemporal dynamics of seizure activity, DeepSIF promises to help guide clinical decisions and improve treatment outcomes for epilepsy patients.