Abstract
In this study, the neuronal current in the brain is represented using Helmholtz decomposition. It was shown in earlier work that data obtained via electroencephalography (EEG) are affected only by the irrotational component of the current. The irrotational component is denoted by $\Psi$ and has support in the cerebrum. This inverse problem is severely ill-posed and requires that additional constraints are imposed. Here, we impose the requirement of the minimization of the $L_2$ norm of the current (energy). The function $\Psi$ is expanded in terms of inverse multiquadric radial basis functions (RBF) on a uniform Cartesian grid inside the cerebrum. The minimal energy constraint in conjunction with the RBF parametrization of $\Psi$ results in a Tikhonov regularized solution of $\Psi$. The RBF shape parameter (regularization parameter), is computed by solving a 1-D non-linear maximization problem. Reconstructions are presented using synthetic data with a signal to noise ratio (SNR) of 20 dB. The root mean square error (RMSE) between the exact and the reconstructed $\Psi$ is RMSE=0.1122. The proposed reconstruction algorithm is computationally efficient and can be vectorized in MATLAB.
Introduction
Electroencephalography (EEG) remains a cornerstone of modern clinical diagnostics, yet it faces a significant challenge: it measures electrical potentials on the scalp's surface, whereas the primary interest lies in the underlying neuronal activity within the brain. Determining the internal current from surface data is a "severely ill-posed" inverse problem, where numerous internal activity patterns can produce identical surface readings. To address this, researchers must apply constraints, such as assuming a "minimal energy" path for brain activity. Traditional methods often struggle with computational intensity or rely on oversimplified spherical models that do not accurately represent human anatomy.
Methodology
The researchers developed a hybrid approach that integrates analytical physics, radial basis functions (RBFs), and machine learning to reconstruct neuronal currents with high precision and efficiency.
Helmholtz Decomposition and RBFs
The core of the method lies in mathematically splitting the neuronal current into two parts. Leveraging the physical principle that EEG data is only influenced by the irrotational component ($\Psi$), the algorithm focuses exclusively on this variable. This component is represented using "inverse multiquadric" RBFs on a Cartesian grid, allowing for a smooth, continuous approximation of activity within the cerebrum.
Head Modeling and Conductivity
To simulate the electrical environment of the head, the study utilizes a multi-layer model. This model accounts for the distinct electrical properties of various tissues, which influence how signals propagate from the brain to the scalp sensors.
The accuracy of the forward model depends heavily on the conductivity values assigned to each domain, such as the cerebrum, cerebrospinal fluid (CSF), skull, and scalp.
Surrogate Modeling and Optimization
Handling complex head geometries typically requires heavy computational resources. To bypass this, the authors employed a surrogate model—a two-layer neural network—to approximate the relationship between internal currents and surface potentials. This allows for rapid calculation without real-time PDE solving. Furthermore, the reconstruction incorporates Tikhonov regularization to enforce the minimal energy constraint, optimizing the RBF "shape parameter" through a 1-D non-linear maximization process.
Results and Discussion
The algorithm was validated using synthetic data from the OpenMEEG package, ensuring robustness by avoiding "inverse crimes" (using different models for data generation and reconstruction).
Key performance metrics include:
- High Fidelity: The reconstruction of the irrotational component $\Psi$ achieved a Root Mean Square Error (RMSE) of 0.1122 at a signal-to-noise ratio (SNR) of 20 dB.
- Surrogate Accuracy: The machine learning surrogate matched high-fidelity boundary element methods closely, with a Relative Mean Distance (RDM) of 0.2102.
- Computational Efficiency: The combination of RBFs and surrogate modeling enables full vectorization in MATLAB, making the process significantly faster than traditional iterative solvers.
Conclusion and Future Work
This hybrid algorithm demonstrates a powerful way to solve the EEG inverse problem by combining physical insights with efficient numerical techniques. While currently validated on synthetic data, future research will focus on applying the model to real-world human EEG data, incorporating temporal dynamics, and further refining the anatomical layers of the head model to improve clinical utility.
Summary updated