Abstract
Analyzing brain imaging data like Magnetoencephalography (MEG) is challenging due to its high dimensionality and irregular spatial distribution. This paper proposes a novel deep learning architecture that treats brain sensors as nodes in a connectivity graph. By combining Graph Convolutional Networks (GCNs) with an autoencoder structure, the model extracts discriminative features in an unsupervised manner. Experimental results on MEG datasets demonstrate that this graph-informed approach significantly outperforms traditional dimensionality reduction techniques in classifying mental states.
Introduction
Traditional brain imaging analysis often struggles with the spatiotemporal complexity and non-Gaussian noise inherent in signals like MEG or EEG. A primary hurdle is that these sensors do not reside on a regular grid, making standard Convolutional Neural Networks (CNNs) unsuitable. Instead, brain activity is better represented as signals on a graph, where nodes are sensors and edges represent anatomical or functional connectivity.
This research addresses the need for dimensionality reduction techniques that respect the brain's underlying "graph-like" nature. By leveraging Graph Signal Processing (GSP), the authors aim to transform high-dimensional, noisy sensor data into a compact, informative representation for downstream tasks like classification.
Proposed Methodology
The core of the proposed system is an autoencoder-like neural network integrated with graph convolutional layers. The architecture follows a two-stage process:
- Graph Convolutional Layers: The model employs spectral graph convolutions to aggregate information from neighboring sensors. To maintain computational efficiency, a first-order Chebyshev polynomial approximation is used. This allows the network to learn localized filters that are sensitive to the brain's connectivity patterns.
- Bottleneck Autoencoder: The high-dimensional outputs from the GCN layers are flattened and passed through a series of fully connected layers. The network is trained end-to-end by minimizing the Mean Square Error (MSE) between the input signal and its reconstruction, ensuring the bottleneck layer captures the most essential features.
Once the unsupervised training is complete, the 50-dimensional feature vector from the bottleneck is used as input for a linear Support Vector Machine (SVM) to classify the subject's mental state.
Experimental Evaluation
The model was evaluated using real MEG signals from three subjects (A, B, and C) who were presented with visual stimuli consisting of human faces and various household objects. Data was collected from 306 sensors, focusing on the 96ms to 110ms post-stimulus window. The underlying graph structure was estimated using Granger Causality Connectivity (GCC) analysis.
Performance Comparison
The proposed model was compared against several baseline methods, including Principal Component Analysis (PCA), Robust PCA (RPCA), Graph Basis Functions (GBF), and standard 2-layer and 4-layer Stacked Autoencoders (SAE).
As shown in the table above, the proposed model consistently achieved the highest classification accuracy across all subjects. For Subject A, the accuracy reached 68.33%, a notable improvement over the 64.82% achieved using raw data and the 66.93% from a standard 4-layer autoencoder.
Discussion and Key Insights
A critical finding of this study is the importance of the graph prior. When the GCC-based graph was replaced with a random or identity matrix, performance dropped significantly, confirming that the specific biological connectivity of the brain is essential for effective feature extraction. Furthermore, the success of unsupervised feature discovery suggests that the model captures intrinsic properties of the brain's response to visual stimuli.
Conclusion and Future Work
This paper demonstrates that integrating graph convolutional layers into an autoencoder framework provides a powerful tool for brain imaging analysis. By respecting the irregular geometry of sensor placement and the functional connectivity of the brain, the model extracts more discriminative features than grid-based or purely statistical methods.
Future research will focus on learning the graph structure and network parameters simultaneously, potentially improving robustness to errors in initial connectivity estimation. Additionally, validating the model on larger datasets will be crucial for its application in clinical diagnostics and real-time brain-computer interfaces.
Summary updated