Sequential Monte Carlo Graph Convolutional Network for Dynamic Brain Connectivity

Ercan Engin Kuruoglu, PhD Cantab, Prof, Ieee Fellow, Fengfan Zhao, Ercan Engin KuruogluICASSP 2024 · DOI

TL;DR

This paper introduces the Sequential Monte Carlo Graph Convolutional Network (SMC-GCN), a novel method for analyzing dynamic brain connectivity by combining particle filtering with graph convolutional networks. This approach overcomes limitations of static connectivity analysis and improves the accuracy of brain disorder classification.

What Problem Does This Paper Address?

Functional connectivity analysis is crucial for understanding brain function, but traditional methods often assume static connections and noise-free data. This is a significant limitation because brain connectivity is inherently dynamic, changing over time. Existing graph-based analyses, while powerful, typically rely on static connectivity matrices derived from the entire fMRI scan, failing to capture these temporal fluctuations. This paper addresses the need for a method that can effectively model and analyze dynamic brain connectivity, particularly in the context of identifying biomarkers for brain disorders.

Methods

The core innovation is the Sequential Monte Carlo Graph Convolutional Network (SMC-GCN). This method extends classical particle filtering, a technique proven effective for tracking problems with noisy observations, to dynamic graph analysis.

Here's a breakdown of the key components:

  • Particle Filtering: This framework is used to estimate the hidden states of a dynamic system (the brain's connectivity) from partial and noisy observations (fMRI data). Instead of traditional vector states, SMC-GCN uses "graph particles," where each particle represents a dynamic graph.
  • Graph Convolutional Network (GCN): A GCN is integrated into the particle filtering framework to model the state transitions. This allows the model to learn complex, non-linear relationships within the graph structure and helps in limiting spurious connections. The GCN acts as the "state evolution model" for the graph particles.
  • Graph Construction: For each time window of the fMRI data, a graph is constructed. Nodes represent brain regions (ROIs), and edges are initially weighted by partial correlations. Feature embeddings for each ROI are also computed.
  • State Transition: The GCN takes a graph particle from the previous time step and predicts the state (graph structure) for the current time step. This is where the dynamic nature of connectivity is modeled.
  • State Update: The predicted graph state is then updated based on the current observation (the graph constructed from the current fMRI time window). A discriminative function is used to up/down-weight the state based on how well it matches the observation.
  • Soft Resampling: A differentiable soft resampling strategy is employed to avoid particle degeneracy while maintaining differentiability for end-to-end training.
  • Graph Particle Classification: Finally, a global pooling operation (mean and max pooling) is applied to the output graph embeddings from the GCN, followed by an MLP for classification.

Datasets

The primary dataset used for evaluation is the Autism Brain Imaging Data Exchange (ABIDE) database. This is a large, challenging database for binary classification tasks, aggregating functional MRI and phenotypic data from 1112 subjects across 20 international sites. The authors used a subset of 578 samples after preprocessing and quality control, with a class distribution of 287 subjects with Autism Spectrum Disorder (ASD) and 291 healthy controls (HC). The data was split using a stratified 5-fold cross-validation approach.

Key Insights and Ideas

  • Dynamic Connectivity is Key: The paper strongly emphasizes that brain connectivity is not static and that capturing these temporal dynamics is crucial for accurate analysis, especially for understanding brain disorders.
  • Particle Filtering for Graphs: Extending particle filtering, a robust technique for sequential data, to handle graph-structured data (graph particles) is a significant methodological advancement.
  • GCN for Non-linear Graph Dynamics: Integrating GCNs into the state transition model of particle filtering allows for the capture of complex, non-linear relationships in dynamic brain connectivity, which linear models would miss.
  • Improved Classification: The SMC-GCN approach demonstrates superior performance in classifying brain disorders compared to existing static graph-based and machine learning methods.

Key Results

The SMC-GCN model achieved superior performance in brain disorder classification on the ABIDE dataset compared to baseline models. Specifically, in the 5-fold cross-validation results presented in Table 1:

  • SMC-GCN achieved an accuracy of 73.64% (±0.98), outperforming SVM (61.87%), RF (60.58%), GCN (68.28%), GAT (66.24%), and BrainGNN (70.66%).
  • It also showed improvements in sensitivity (70.06% vs. others) and AUC (75.29% vs. others).
  • The paper also includes an ablation study on the number of particles, showing a smooth growth in accuracy and AUC as the number of particles increases from 10 to 50, indicating the model's robustness to particle count within a reasonable range.

Limitations and Future Work

  • Particle Number Optimization: While the ablation study shows performance improves with more particles, the authors note that the training time doubles with each increase. Further investigation into the efficiency and optimal number of particles for sampling approximation is suggested.
  • Generalizability of Graph Particles: The paper frames its contribution as extending particle filtering to "particles could correspond to graphs or other data embedding structures." Future work could explore applying SMC-GCN to other types of dynamic graph topologies beyond brain connectivity.
  • Other Applications: The authors express interest in exploring other applications of the SMC-GCN model and its capacity for handling a growing number of particles.

Summary updated

BibTeX
@inproceedings{fellow2024sequential,
  author = {Ercan Engin Kuruoglu, PhD Cantab, Prof, IEEE Fellow and Fengfan Zhao and Ercan Engin Kuruoglu},
  title = {{Sequential Monte Carlo Graph Convolutional Network for Dynamic Brain Connectivity}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2024},
  doi = {10.1109/icassp48485.2024.10448129},
}
Figure