Research On Robust EEG Extended Source Imaging Algorithm Based On L1-Norm Residual

Reconstructing the locations and extents of cortical activities accurately from EEG signals plays an important role in neuroscience research and clinical surgeries.However,due to the ill-posed nature of EEG source imaging problem,there exist no unique solutions for the measured EEG signals.Therefore,some classical imaging algorithms use L2 norm regularization or L1 norm regularization to obtain the unique solution of the inverse problem.Additionally,evoked EEG is inevitably contaminated by strong background activities and outliers caused by ocular or head movements during recordings.In order to deal with these outliers to improve the performance of the imaging methods,L₁norm residuals and variation sparsity are used to achieve more robust reconstruction of brain source locations and extents.The main innovations of this thesis are as follows:1.A new method is proposed,L1R-VSSI(L1-norm Residual Variation Sparse Source Imaging).L1R-VSSI employs the L1-loss for the residual error to alleviate the outliers.Additionally,the L1-norm constraint in the variation domain of sources is implemented to achieve globally sparse and locally smooth solutions.The solutions of L1R-VSSI is obtained by the alternating direction method of multipliers(ADMM)algorithm.Results on both simulated and experimental EEG data sets show that L1R-VSSI can effectively alleviate the influences from the outliers during the recording procedure.L1R-VSSI also achieves better performance than traditional L2-norm based methods(e.g.,w MNE,LORETA)and sparse constraint methods in the original domain(e.g.,SBL)and in the variation domain(e.g.,VB-SCCD).2.A new method is proposed,L1R-SSSI(L1-norm Residual Structured Sparsity Source Imaging).L1R-SSSI employs the L1-loss to model the residual error and adopts the structured sparsity constraint,which incorporates the L1-norm regularization in both the variation and original source domain.The sparsity constraint on the original source domain is further added to limit the global energy to obtain a more accurate amplitude description.The estimations of L1R-SSSI are efficiently obtained using the ADMM algorithm.The experimental results show that L1R-SSSI outperforms the traditional L2-norm based methods,and SISSY,which employs L2-norm loss and structured sparsity.L1R-SSSI also provides better estimations of extended sources than the method with L1-loss and Lp-norm regularization term(e.g.,LAPPS).3.The performance of the imaging algorithm is verified by using an EEG data set containing accurate known positions in the brain.Using such a"truth"EEG data to compare these algorithms.The experimental results show that the position error of the stimulation points located by L1R-SSSI method is the smallest,and there is no obvious pseudo source.