Brain-imaging based methodology for OPM sensor placement

Abstract

ABSTRACT: Optically-pumped magnetometers (OPMs) have reached sensitivity levels that make them viable portable alternatives to traditional superconducting technology for magnetoencephalography. OPMs do not require cryogenic cooling, and can therefore be placed directly on the scalp surface. Unlike cryogenic systems based on a well characterised xed arrays essentially linear in applied ux, or electroencephalography sensors that do not need to account for sensors orientation; OPM sensors are no longer rigidly arranged with a scanner system. Therefore, uncertainty in their locations and orientations with respect to the brain, and with respect to one another, must be accounted for. In this thesis dissertation, we propose a methodology to estimate the true sensor geometry of a disturbed array. We use parametric Bayesian inversion methods to perform neural source reconstruction and score among disturbed geometries with Free Energy as a cost function. This geometry disturbance is non-linear, causing local sub-optimal values on Free Energy that we tackle with a Metropolis search. Looking for a robust solution to this sensor placement problem, we develop a Multiple Kernel Learning (MKL) approach to extract the predominant complex dynamics hidden in the data. To do this, a weighted mixture of Gaussian kernels is used to highlight the data relationships, enhancing the data-driven covariance estimation and leading to a more reliable neural source reconstruction. When tested over disturbed OPM geometries, the MKL based solvers turned the Free Energy into a monotonic function, allowing the use of gradient descent optimisation. As a result, we estimate the true geometry of disturbed OPM arrays with a similar error than Metropolis search, but with 90% fewer iterations and allowing a larger search space. Our proposal suggests that a exible and scalable design for sensor placement can be used to harness the potential of OPMs.