Study On Cortical Source Localization Method For Multi-channel Electroencephalogram

Electroencephalogram(EEG)is the potential signal attributed to electrical activities accompanying activation of cortical neurons,which can be collected by electrodes.Scalp EEG has the advantages of non-invasiveness,easy implementation,high temporal resolution,and low cost.It is an important tool for brain function cognition,disease diagnosis,and brain-computer interface research.The analysis objects of EEG signals include transient waveform and oscillatory components,corresponding to the real temporal features and the frequency-domain complex features,respectively.However,the EEG signals recorded by the scalp electrodes are only the results of the superposition of all activated sources in the cerebral cortex after volume conduction,so they cannot accurately reflect the true source activation at the corresponding cortical locations.The inverse problem of EEG aims at making inference of source signals in the cortex from the EEG signals recorded by multi-channel scalp electrodes.This problem is a severe ill-posed one.Although the existing methods tried to explore the plausible priors related to the source signal distributions on the cortex,there are still problems to be solved.These problems include: inappropriate way to handle the noises,cortical functional partition and non-negativity of variance parameters in each region being not considered,no rigorous signal model for oscillatory EEG frequency-domain components,etc.With the above problems addressed,the contributions of this dissertation are given as follows.The first work in this dissertation is to study the cortical source localization method for transient EEG waveform.As electrical activities are more likely to be generated by neural masses located at various functional brain regions on the cortex,in recent years block-sparsity priors of cortical sources have been exploited to develop the cortical source localization methods,where the block-sparsity defines both intra-block source correlation and inter-block sparsity at source level.In this dissertation,the block-sparsity of sources distributed on the cortex surface was utilized to propose a brain source localization method for transient EEG waveform in the framework of sparse Bayesian learning(SBL).There are two main contributions in the proposed method.First,in building hierarchical probabilistic generative model for the block-sparse signal recovery problem,the distribution of sample covariance matrix of sensor measurements instead of the data itself is considered.Such an alteration at sensor-level modeling makes the requirement of noise covariance matrix be circumvented,while as aforementioned this matrix should be known or estimated in the existing SBL based algorithms.Second,after a series of transformations,the block-sparse signal recovery problem is transformed to a conventional atom-sparse one,where variance parameters of brain regions are only needed to be estimated.Considering the nonnegativity and sparsity of variance parameter vector,nonnegative Gaussian distribution was introduced as its prior to enhance the source localization performance.Simulations demonstrated that the proposed nonnegative block-SBL(NNBSBL)algorithm achieved superior performances in cortical source detection and localization,compared to benchmark and state-of-the-art algorithms.The performance of the proposed algorithm is also evaluated through real P300 EEG data,which is proved consistent with the conclusions of P300 source locations in literature.The second work in this dissertation is to study the cortical source localization method for oscillatory EEG at frequency domain,where two consecutive studies have been carried out.First,the signal model for frequency-domain components of oscillatory EEG was given,by proving that the multi-channel EEG Fourier components at the interested frequency equal multiplying the lead field matrix by a complex-valued vector,contaminated by spontaneous EEG and electrical noise.Based on this model,steming from the atom-sparsity prior for cortical source distribution,a common SBL algorithm for localizing frequency-domain components of oscillatory EEG has been proposed.By the proposed algorithm,not only a single subject’s brain source localization can be achieved,but also the common locations of a certain type of oscillatory EEG integrating multiple subjects can be given,even when the electrode layout or number of electrodes varies among subjects.Second,the block-sparsity prior was further considered,and how to apply the real-valued NNBSBL algorithm designed for localizing cortical sources of transient EEG waveform to this complex-valued problem was studied.The proposed solution is based on the distribution of the sample covariance matrix of the complex component vector at the interested frequency,which is generalized complex Gaussian distribution.After a series of transformations and constructions performed on this sample covariance matrix,a real-valued signal model similar to that in the transient waveform cortical source localization problem was obtained,based on which the NNBSBL algorithm was applied.Both simulation experiments and verification experiments on real EEG data have proved the superior performance of the proposed method compared to the existing methods.The studies in this dissertation addressed the problems of localizing cortical sources for transient EEG waveforms as well as frequency-domain components of oscillatory EEG.Steming from rigorous models,robust cortical source localization methods have been built based on the idea of SBL,making full use of neurophysiological features that are closer to reality.The proposed brain source localization methods are expected to provide more powerful tools for future brain research.