PDE-constrained optimization for electroencephalographic source reconstruction

TL;DR

This paper presents a new finite-element based numerical method for EEG source reconstruction, formulating it as an optimal control problem that directly incorporates noisy electrode data, avoiding lead-field matrix formation.

Contribution

It extends the mixed quasi-reversibility method by explicitly including noisy data and discretizes the problem in finite-element spaces, enabling handling complex MRI-based meshes.

Findings

Accurately reconstructs cortical activity in tests.

Handles MRI-based meshes of arbitrary complexity.

Does not require lead-field matrix formation.

Abstract

This paper introduces a novel numerical method for the inverse problem of electroencephalography(EEG). We pose the inverse EEG problem as an optimal control (OC) problem for Poisson's equation. The optimality conditions lead to a variational system of differential equations. It is discretized directly in finite-element spaces leading to a system of linear equations with a sparse Karush-Kuhn-Tucker matrix. The method uses finite-element discretization and thus can handle MRI-based meshes of almost arbitrary complexity. It extends the well-known mixed quasi-reversibility method (mQRM) in that pointwise noisy data explicitly appear in the formulation making unnecessary tedious interpolation of the noisy data from the electrodes to the scalp surface. The resulting algebraic problem differs considerably from that obtained in the mixed quasi-reversibility, but only slightly larger. An interesting feature of the algorithm is that it does not require the formation of the lead-field matrix. Our tests, both with spherical and MRI-based meshes, demonstrates that the method accurately reconstructs cortical activity.