Brain Source Localization in the Presence of Leadfield Perturbations

Abstract

Brain source localization enables us to localize different areas of the brain that are activated during any mental activity. This thesis makes use of Electroencephalography (EEG) recordings which is an important noninvasive tool for studying the temporal dynamics of the human brain. EEG source localization finds its applications in cognitive neuroscience in order to develop a Brain Computer Interface (BCI), and in psychopharmacology and psychiatry, to localize sources in certain frequency bands. Unfortunately, EEG readings cannot directly indicate the location of the source of brain activity using the signals measured on the scalp, which contributes to the ambiguity of the inverse problem. In order to solve the ill-posed inverse problem, array processing methods are implemented, in conjunction with various techniques that are applied, to improve the localization in the presence of calibration errors. In this thesis, a recently developed G-MUSIC algorithm is applied to the problem of brain source localization. G-MUSIC is a form of weighted MUSIC that performs better in scenarios where only limited sample support is available. Two transfer function based calibration algorithms are also developed to estimate the accurate location of neural activity in the brain when the measured leadfield is perturbed. The localization performance of G-MUSIC is compared to the traditional MUSIC algorithm and quantified in terms of the localization error. This thesis also addresses the problem of localization when exact knowledge of the leadfield matrix, for an individual head anatomy, is not available, by developing an iterative algorithm. This algorithm includes a high resolution localization technique, recently used in radar field, called Source Affine Image Reconstruction Algorithm (SAFFIRE) that can determine the model order (number of sources) and their locations. A beamformer is then designed in order to estimate the dipole source amplitudes. Finally, the EEG signal is reconstructed and related to the actual EEG signal via a calibration matrix. This procedure is repeated until a convergence criteria is met. The performance of this algorithm is quantified in terms of the localization error and accuracy and further validated by applying it to experimental data. In conclusion, the algorithm is also tested on non-stationary EEG signal, where a variant of the conventional adaptive beamformer is applied in order to estimate the source signal amplitudes.