Abstract
Electrophysiological source imaging (ESI) refers to the process of reconstructing underlying activated sources on the cortex given the brain signal measured by Electroencephalography (EEG) or Magnetoencephalography (MEG). Due to the ill-posed nature of ESI, solving ESI requires the design of neurophysiologically plausible regularization or priors to guarantee a unique solution. Recovering focally extended sources is more challenging, and traditionally uses a total variation regularization to promote spatial continuity of the activated sources. In this paper, we propose to use graph Fourier transform (GFT) based bidirectional long-short term memory (BiLSTM) neural network to solve the ESI problem. The GFT delineates the 3D source space into spatially high, medium and low frequency subspaces spanned by corresponding eigenvectors. The low frequency components can naturally serve as a spatially low-band pass filter to reconstruct extended areas of source activation. The BiLSTM is adopted to learn the mapping relationship between the projection of low-frequency graph space and the recorded EEG. Numerical results show the proposed GFT-BiLSTM outperforms other benchmark algorithms in synthetic data under varied signal-to-noise ratios (SNRs). Real data experiments also demonstrate its capability of localizing the epileptogenic zone of epilepsy patients with good accuracy.
Introduction
Electrophysiological source imaging (ESI) is a critical technique in neuroscience for reconstructing cortical activation from non-invasive measurements like Electroencephalography (EEG). However, ESI is inherently an "inverse problem" and is fundamentally ill-posed: there are thousands of potential source locations in the 3D volume of the brain, but only a limited number of sensors on the scalp. Traditional methods often struggle with "focally extended sources," where neural activity is not a single point but a small, continuous region. Standard algorithms frequently produce "blurry" results or fail to capture the spatial continuity of these active regions.
Methodology
The proposed GFT-BiLSTM framework addresses these challenges by integrating spatial graph theory with temporal deep learning to map EEG signals back to their neural sources.
Graph Fourier Transform (GFT) for Spatial Regularization
The researchers treat the 3D brain source space as a graph, where vertices represent voxels and edges represent physical proximity. By computing the Laplacian matrix of this graph, they apply a Graph Fourier Transform (GFT) to decompose the brain's complex spatial activity into "frequencies." Just as a standard Fourier transform separates a sound into high and low pitches, the GFT separates brain activity into high and low spatial frequencies.
Because extended brain activity is spatially "smooth," it can be represented almost entirely by low-frequency components. The authors project the thousands of potential sources into a much smaller subspace of low-frequency eigenvectors. This serves as a natural spatial low-pass filter, effectively baking neurophysiological priors into the architecture and drastically reducing the number of parameters the neural network needs to learn.
BiLSTM Mapping and Reconstruction
A Bidirectional Long Short-Term Memory (BiLSTM) network is trained to take the time-varying EEG signal as input and predict the low-frequency GFT coefficients. Because it is bidirectional, the network utilizes the entire temporal evolution of a brain signal to determine its origin. Finally, an Inverse GFT (IGFT) is applied to the predicted coefficients to map the signal back into the 3D brain space.
Experimental Results
The model was evaluated using both synthetic simulations and real-world clinical data.
Synthetic Evaluation
In a simulated environment with 2,052 brain sources and 128 electrodes, GFT-BiLSTM significantly outperformed benchmark algorithms like dSPM, MNE, and sLORETA. In tests at a signal-to-noise ratio (SNR) of 20dB, GFT-BiLSTM achieved a localization error (LE) of 15.57, while the next best traditional method (dSPM) lagged at 58.18. The proposed model maintained an Area Under Curve (AUC) above 0.96 even in noisy conditions, indicating high precision in identifying the boundaries of active regions.
Clinical Application
In a real-world epilepsy case involving 29-channel EEG recordings, GFT-BiLSTM localized the epileptogenic zone to the left frontal region with much tighter focus and better spatial continuity than traditional methods. This localization matched the area that was surgically removed to successfully treat the patient, providing strong clinical validation.
Conclusion
GFT-BiLSTM offers a robust solution for ESI by leveraging the spatial priors inherent in graph-based frequency analysis and the temporal modeling capabilities of BiLSTMs. While the choice of the low-frequency threshold remains empirical, the model demonstrates significant improvements in both accuracy and spatial resolution for reconstructing extended neural sources.
Summary updated