Department of Computer, Control and Management Engineering

Electroencephalography (EEG) provides millisecond-scale resolution of neural activity but struggles to accurately localize multiple concurrent sources, especially in spatially close regions. Classical linear inverse methods, such as MNE, sLORETA, and dSPM, address the ill-posed inverse problem through regularization but often exhibit a “single-source bias”, suppressing smaller generators. This paper introduces a deep learning framework designed to robustly identify multiple sources of activity from short EEG segments. Our approach leverages a realistic simulation pipeline that systematically generates EEG recordings from physiologically plausible, distributed current sources. We train a convolutional neural network (ConvNET) on thousands of such simulations, ensuring generalization by using a forward model distinct from that of classical solvers, thereby minimizing the risk of an "inverse crime". We evaluate our ConvNet against nine well-established inverse solvers (MNE, dSPM, sLORETA, eLORETA, LORETA, LAURA, and depth-weighted variants). Benchmarking across multiple synthetic test scenarios demonstrates that our method consistently outperforms traditional solvers, particularly in resolving closely spaced sources, while maintaining or improving accuracy for single-source cases. These results highlight the potential of deep learning to overcome biases in EEG source imaging, offering a more reliable approach for multi-source localization.