Model based deep learning and graph neural networks for EEG source imaging

Résumé fourni par la source

EEG source imaging (ESI) aims to estimate the activity of neuronal sources in the brain based on the electroencephalography (EEG) signal. This is an ill-posed inverse problem that requires the addition of prior information to find a solution. Choosing an appropriate prior for ESI remains challenging, particularly since brain activity is a complex signal. In this context, alongside classical optimization approaches used to tackle such inverse problems, deep learning methods have also been explored to learn a direct inverse mapping from EEG to source activity. While such approaches can demonstrate strong performance, they often lack interpretability and adaptability to new head models or subjects, requiring retraining for each new subject configuration. In this work, we study a modelbased deep learning approach that combines physical forward information of EEG with learnable components to obtain a more flexible and interpretable model. The proposed method is a bi-level optimization approach with a learned gradient-based solver for the lower-level cost and a learnable regularization term to better capture the distribution of the observations. Both the regularizer and the optimizer are parametrized by neural networks working in the source domain. To exploit the brain's geometric structure and spatial dependencies, both networks are implemented using graph neural networks (GNNs) defined on the brain mesh. This design allows the model to capture spatial information on the sources more effectively and to be more flexible to changes in the geometry. We evaluate the method against two non-learning-based ESI methods and two direct inversion deep learning models. Results shows that the use of GNNs enables the spatial characteristics of the data to be captured more effectively than with a linear spatial representation. This framework can generalize to different electrode montages, and GNNs demonstrate superior generalization capabilities compared to convolution-based approaches. The model-based deep learning method associated with graph neural networks presented in this work provides a more interpretable and flexible framework for ESI compared to purely data driven methods.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.