Electroencephalography (EEG) offers high temporal resolution but struggles to accurately localize subcortical activity, partly due to the ill-posed nature of the inverse problem and the weak signals from deep structures. Traditional regularized inverse methods are computationally efficient yet often miss deep sources. Here, we introduce a deep learning pipeline specifically designed for subcortical EEG source localization. We generate realistic training data through a custom simulator that combines spatially structured dipole activity, autoregressive time series, controlled synchronization, and distinct forward operators to reduce the inverse crime. Our network maps raw EEG segments directly to subcortical activation, bypassing explicit dipole reconstructions. Compared against nine classical solvers (including MNE, dSPM, sLORETA) across seven different metrics, our approach demonstrates superior localization accuracy and spatial specificity in both cortical and subcortical tests. This mitigates the surface bias typical of standard solutions and highlights the potential of end-to-end deep learning for EEG-based subcortical neuroimaging. Future work will refine simulation realism, explore multi-subject adaptability, and address transfer to real EEG.