Improving EEG Source Localization with a Novel Regularization: Spatiotemporal Graph Total Variation (STGTV) Method.

Abstract

Electroencephalography (EEG) source Iocalization aims at reconstructing the current density on the brain cortex from scalp EEG recordings. It of ten starts with a generative model that maps brain activity to the EEG recording, and then solves the inverse problem. Previously proposed method graph fractional-order total variation (gFOTV) is based on spatial regularization, and was shown superior to some other existing spatial-regularized methods in simulation tests. However, the gFOTV addresses inverse problem for one time point at a time. The resultant estimated times series of brain activity is a simple concatenation of reconstructions independently performed at each time instance, and risks spurious temporal discontinuity due to overfitting noise in EEG recordings. In addition, the performance is subject to low signal-to-noise ratio (SNR) and small number of electrodes, which happens in realistic EEG recordings. To account for the generally continuous temporal variation in brain activity, but also allow for properly triggering abrupt changes, we propose a novel formulation that incorporates spatiotemporal regularization. Specifically, our method, called spatiotemporal graph total variation (STGTV) adopts graph fractional-order total variation (gFOTV) for spatial regularization and total variation (TV) for temporal regularization. The gFOTV encourages spatially smooth source distributions, and the temporal TV enhances temporal consistency in estimated activity maps. The introduction of implicit temporal coupling by temporal TV also helps with noise cancelation and enhances SNR. In a simulation study, the performance of the proposed method was compared against that from the gFOTV regularization alone. The results showed that the proposed STGTV method significantly improved gFOTV, with lower Iocalization errors and less spuriously discovered sources.