Developing and Validating Realistic Head Models for Forward Calculation of Electromagnetic Fields with Applications in EEG

Localizing the sources responsible for generating an observed electroencephalography (EEG) signal can be useful in several contexts. For example, it may help guide presurgical planning in epilepsy or improve connectivity estimates obtained from EEG data. However, EEG source analysis is an ill-posed, inverse problem and in order to solve this, we first need to solve the corresponding forward problem. The forward problem is the task of determining the measurements corresponding to individual neural sources. The accuracy with which we can solve this problem is determined (in part) by our ability to construct a valid volume conductor model. Constructing such a model includes modeling of the anatomy of the human head as well as the conductive properties of different tissue compartments. We also need to specify the positions of the electrodes relative to the head and choose a suitable resolution for our model. In this thesis we seek to investigate how the different aspects of forward modeling affect not only the forward solution itself but also the corresponding inverse solution. We focus on the anatomical accuracy of the volume conductor model as well as the accuracy with which electrode positions are specified. As to the inverse solution, we will primarily be concerned with assessing localization accuracy (as opposed to, for example, determining the strength of a sourceactivation). First, we present a pipeline (available in SimNIBS) for generating reasonably realistic anatomical models of the human head with particular emphasis on reconstructing the skull. This pipeline is based on SPM12 and CAT12 where the latter is used to improve the accuracy of the brain tissue segmentation. We compare with existing tools from FSL and BrainSuite and show that the new pipeline improves skull segmentation, particularly when using both T1-and T2-weighted magnetic resonance imaging (MRI) scans. We also show that it is important to ensure that the structural images being used are of high quality (e.g., by minimizing artifacts such as fat shift) in order to facilitategood segmentation results. Subsequently, we compare forward solutions generated by SimNIBS, MNEPython, and FieldTrip. The major difference between these models is the extent to which they are able to capture the underlying anatomy with SimNIBS generally being more accurate than the latter two. We find increased topographic and magnitude errors of the forward solutions from MNE-Python and FieldTrip compared to SimNIBS throughout most of the brain suggesting large overall differences in the forward solutions. We also compare with a model based on a template anatomy. This too, resulted in substantial errors. In addition to comparing different pipelines, we also compare different ways of specifying the electrode positions. In particular, we compared digitizing the electrodes against using a template description of the electrode positions which is adapted to each subject. We investigated two templates; one which we created by digitizing the relevant cap on a 3D printed model of the MNI head and another which used the positions specified by the manufacturer. We found substantial topographic errors when using the manufacturer layout especially in occipital and parietal areas which was also where we found the largest errors in electrode locations (compared to the digitized positions). Our custom template performed better in these areas suggesting that the way in which the template positions have been generated can affect accuracy significantly In a follow-up study, we investigate the feasibility of optimizing the electrode positions obtained using our custom template based on a few measurements of distances and angles between electrodes and landmarks. We used measurements between nasion, left preauricular (LPA), right preauricular (RPA), inion, and nearby electrodes for a total of the eight measurements (four distances and four angles). We show that the result of the optimization is not particularly affected by errors in the measurements but that the effect across subjects differ substantially. Specifically, some subjects benefitted considerably whereas it didnot make much of a difference for others suggesting that the procedure is good at preventing outliers. Finally, we explore the effect of forward solution errors on source localization errors. Using the same forward models described above, we simulated data at different signal-to-noise ratio (SNR) levels and used different inverse methods(dynamic statistical parametric mapping (dSPM), standardized low resolution electromagnetic tomography (sLORETA), dipole fitting, multiple signal classification (MUSIC)) for localization. In general, we found that dipole fitting and MUSIC were slightly more sensitive to errors in the forward model compared to the minimum norm estimate (MNE)-based ones whereas all methods were sensitive to SNR level. Using a template anatomy resulted in larger errors compared to anatomical models generated in SimNIBS, MNE-Python, or FieldTrip. On the other hand, using a template anatomy with digitized electrode positions performed better than using the correct anatomical model for each subject but with the manufacturer description of electrode positions. This highlights the importance of correctly specifying electrode positions as otherwise large, spatially correlated errors may be induced.