Abstract
Electroencephalography source imaging (ESI) is an ill-posed inverse problem: an additional constraint is needed to find a unique solution. The choice of this constraint, or prior, remains a challenge for most ESI methods. This work explores the application of supervised learning methods for spatio-temporal ESI, where the relationship between measurements and sources is learned directly from the data. Three neural networks were trained on synthetic data and compared with non-learning based methods. Two distinct types of simulation, each based on different models of brain electrical activity, were employed to quantitatively assess the generalization capabilities of the neural networks and the impact of training data on their performances, using five complementary metrics. The results demonstrate that, with appropriately designed simulations, neural networks can be competitive with non-learning-based approaches, even when applied to previously unseen data.