Robustness of MEG source imaging in pediatric patients with surgical cavities: A case series

Magnetoencephalography (MEG) is used to localize the epileptogenic zone (EZ) in children with drug-resistant epilepsy (DRE) undergoing pre-surgical evaluation. However, its accuracy may be compromised in patients with cerebrospinal fluid (CSF)-filled surgical cavities from prior failed resections. Finite element method (FEM) head models of tissue geometry and biophysics may better account for altered anatomy than the commonly used spherical approximations, but the added value in real clinical cases remains unclear.

We evaluated the impact of head model choice on MEG source imaging (MSI) in pediatric DRE patients with pre-existing surgical cavities undergoing reoperation.

We retrospectively analysed interictal and ictal events from four patients (ages 9-18) with prior failed epilepsy surgery who underwent MEG before a second or third resection. Dynamic statistical parametric mapping (dSPM) was applied to all six activity datasets. Equivalent current dipole (ECD) estimation retained dipoles for the four interictal datasets; neither ictal dataset yielded dipoles meeting the goodness-of-fit threshold. Source imaging used conventional overlapping-spheres (OS), boundary element method (BEM), advanced 3-layer FEM (3L-FEM), and 4-layer FEM (4L-FEM, including cavities) head models. EZ estimates were compared in terms of volume, EZ-cavity centroid distances, overlap (DICE coefficients), and agreement with the subsequent resection zone (RZ).

Within each mapping approach, distances from dSPM centroids and dipole locations to the cavity differed across head models by less than 1 cm for all but one activity dataset. Pairwise dSPM source-centroid distances reached 1.84 cm, and pairwise dipole distances reached 4.39 cm in one participant's MEG session. OS and BEM dSPM estimates were typically larger in volume than FEM counterparts, with FEM providing more spatially restricted sources. In the seizure-free cases, the dSPM source imaging from 3 out of 4 epileptiform events overlapped with the subsequent RZ. Computationally, OS models were generated within seconds, BEM models in ∼7 min, whereas 4L-FEM required >20 min per head model.

In this pediatric epilepsy case series with surgical cavities, MSI showed broad agreement across simple spherical and advanced realistic head models. These findings did not establish a consistent localization advantage of one head model over another. However, larger prospective cohorts with independent ground truths are needed to validate and generalize these findings across cavity types and EZ locations.