What actually matters in multi-compartment EEG head models: A controlled FEM study of parcellation granularity, skull layering, mesh quality, noise, and inverse solver

Abstract

Electroencephalography (EEG) source localization relies on realistic head volume conductor models, yet the practical benefit of highly detailed multi-compartment head models remains uncertain. In this study, we examine how noise level, anatomical parcellation granularity, skull modeling, and inverse-solver choice jointly shape the fidelity and accuracy of EEG forward and inverse solutions. Six multi-compartment head models, spanning a controlled range of segmentation granularity and skull representations, were generated through a transferable workflow applicable to any individual MRI or CT dataset. Forward solutions were computed using a charge-conserving (H(div)) finite element formulation, while inverse solutions were estimated with sLORETA and Dipole Scan, two established benchmark methods representing distributed and focal EEG source localization, respectively, under realistic signal-to-noise ratio (SNR) conditions. Under these conditions, Dipole Scan was more SNR-sensitive and yielded sharper localization at high SNR, whereas sLORETA was less SNR-sensitive and produced more spatially diffuse estimates. Within an FEM framework with isotropic, piecewise-constant conductivity assumptions, finer subdivision of already detailed FEM head models did not consistently improve source-localization accuracy in the absence of meaningful conductivity contrasts. Localization performance was instead driven primarily by SNR, solver choice, skull representation, mesh quality, and void elimination. Among anatomically comparable models, leadfield power and localization error distributions largely overlapped under matched solver and noise conditions, indicating limited practical benefits can be gain from additional parcellation within this modeling regime. The inverse analysis in this study was designed as a controlled comparison of inverse-method choices rather than as a ranking of algorithms. In the main study, Dipole Scan and sLORETA were used as representative of focal and distributed approaches, respectively, while the appendix extends the analysis with HAL1R and eLORETA. The results indicate that the observed differences cannot be attributed solely to broad inverse-method paradigms, as methods within the same class can still exhibit meaningful differences in error behavior.