Accurate localization of cortical and subcortical sources of M/EEG signals by a convolutional neural network with a realistic head conductivity model: Validation with M/EEG simulation, evoked potentials, and invasive recordings

While electroencephalography (EEG) and magnetoencephalography (MEG) are well-established non-invasive methods in neuroscience and clinical medicine, they suffer from low spatial resolution. Particularly challenging is the accurate localization of subcortical sources of M/EEG, which remains a subject of debate. To address this issue, we propose a four-layered convolutional neural network (4LCNN) designed to precisely locate both cortical and subcortical source activity underlying M/EEG signals. The 4LCNN was trained using a vast dataset generated by forward M/EEG simulations based on a realistic head volume conductor model. The 4LCNN implicitly learns the characteristics of M/EEG and their sources from the training data without need for explicitly formulating and fine-tuning optimal priors, a common challenge in conventional M/EEG source imaging techniques. We evaluated the efficacy of the 4LCNN model on a validation dataset comprising forward M/EEG simulations and two types of real experimental data from humans: 1) somatosensory evoked potentials recorded by EEG, and 2) simultaneous recordings from invasive electrodes implanted in the brain and MEG signals. Our results demonstrate that the 4LCNN provides robust and superior estimation accuracy compared to conventional M/EEG source imaging methods, aligning well with established neuroscience knowledge. Notably, the accuracy of the subcortical regions was as accurate as that of the cortical regions. The 4LCNN method, as a data-driven approach, enables accurate source localization of M/EEG signals, including in subcortical regions, suggesting future contributions to various research endeavors such as contributions to the clinical diagnosis, understanding of the pathophysiology of various neuronal diseases and basic brain functions.