Applying neural networks for improving the MEG inverse solution

Abstract

Magnetoencephalography (MEG) and electroencephalography (EEG) are appealing non-invasive methods for recording brain activity with high temporal resolution. However, locating the brain source currents from recordings picked up by the sensors on the scalp introduces an ill-posed inverse problem. The MEG inverse problem one of the most difficult inverse problems in medical imaging. The current standard in approximating the MEG inverse problem is to use multiple distributed inverse solutions – namely dSPM, sLORETA and L2 MNE – to estimate the source current distribution in the brain. This thesis investigates if these inverse solutions can be "post-processed" by a neural network to provide improved accuracy on source locations.

Recently, deep neural networks have been used to approximate other ill-posed inverse medical imaging problems with accuracy comparable to current state-of- the-art inverse reconstruction algorithms. Neural networks are powerful tools for approximating problems with limited prior knowledge or problems that require high levels of abstraction. In this thesis a special case of a deep convolutional network, the U-Net, is applied to approximate the MEG inverse problem using the standard inverse solutions (dSPM, sLORETA and L2 MNE) as inputs.

The U-Net is capable of learning non-linear relationships between the inputs and producing predictions about the site of single-dipole activation with higher accuracy than the L2 minimum-norm based inverse solutions with the following resolution metrics: dipole localization error (DLE), spatial dispersion (SD) and overall amplitude (OA). The U-Net model is stable and performs better in aforesaid resolution metrics than the inverse solutions with multi-dipole data previously unseen by the U-Net.