A Novel Scanning Algorithm for MEG/EEG imaging using Covariance Partitioning and Noise Learning

In this paper, we present a novel scanning algorithm, called Covariance Optimization Garnering Noise for Active Cancellation (COGNAC), for magnetoencephalography (MEG) and electroencephalography (EEG) source localization. COGNAC uses a probabilistic graphical generative model for describing sensor data. This novel generative model partitions contributions to sensor data from sources at a particular scan location and from sources outside the scan location, with corresponding multi-resolution variance parameters that are estimated from data. Maximizing a convex upper bound on the marginal likelihood of the data under this generative model results in a cost function that can be optimized efficiently. Importantly, this generative model enables learning of sensor noise without the need for additional baseline or pre-stimulus data. The resulting inference algorithm is quite robust to reconstruction of highly correlated sources and to the effect of high levels of interference and noise sources. Algorithm performance was compared to representative benchmark algorithms on both simulated and real brain activity. In simulations, performance of our novel algorithm is consistently superior to benchmarks. We also demonstrate that the new algorithm is robust to correlated brain activity present in real MEG/EEG data.