Selecting volume conductor models for EEG source localization of epileptic spikes: preliminary results based on 4 operated epileptic patients

In order to reconstruct the underlying active neural sources of epileptic spiking activity measured with scalp- EEG, it is necessary to construct a volume conductor model (VCM) reflecting the geometrical and electromagnetic properties of the head. Often three-shell spherical VCMs are used [1]. Based on boundary element methods (BEM) the patient’s anatomy can be approximated better. In these models, the head is modelled as a 3-layered VCM including scalp, skull and brain tissue [2]. The most realistic models can be constructed using finite element or finite difference techniques (FDM) [3]. Using these techniques many extra tissue types can be incorporated such as CSF, grey and white matter based on MRI tissue segmentation techniques. In this study we investigated the effect of the VCM on spike localization. We used a FDM forward solver to compare three types of models: (i) three-layered spherical models, (ii) three-layered BEM-like models, (iii) 5-layered models using SPM segmentation and (iv) 5-layered models using a segmentation method based on level-set (LS) [4]. For each of the models we assumed the same conductivity for equivalent tissue classes and constrained the solution space to the grey matter subspace segmented from the MR [4]. We used a single dipole model to reconstruct 27-channel EEG spiking activity in 4 patients that underwent epilepsy surgery and have been seizure-free since then. CT-images were available to extract the electrode positions. We applied single dipole spike localization using least squares minimisation and validated the estimated sources against the resected area, as indicated by the post-operative MRI and serving as our ground truth. The closest distance of the estimated dipole to the resected area was used as a measure of localization accuracy. In all 4 patients we found the smallest distances using the most realistic 5-layered VCMs (iii) and (iv). The maximum error was 5 mm. We didn’t find significant differences between level-set and SPM segmentation. For the BEM and the spherical models, the error was bigger ranging from 5 mm to 40 mm from the resected zone. We can therefore preliminary conclude that it is worth the effort to build realistic VCMs including more tissue types, based on the MRI segmentation of the patient’s head. This is especially useful in the presurgical evaluation of patients with refractory epilepsy, where the precise determination of the ictal onset zone is very important