Topology-Learnable Static-Dynamic Graph Convolutional Network for Brain Disorder Detection With Functional MRI.

Area of Science:

• Neuroimaging
• Machine Learning
• Graph Theory

Background:

• Graph convolutional networks (GCNs) show promise in brain disorder classification using functional connectivity (FC) graphs from rs-fMRI.
• Existing GCN methods often rely on static FC (sFC) with fixed topologies, potentially limiting feature learning.
• Dynamic FC (dFC) variations are underutilized in current disease recognition models.

Purpose of the Study:

• To propose a novel topology-learnable static-dynamic graph convolution network (TSD-GCN) for automated brain disorder identification.
• To adaptively learn topological structures from both static and dynamic FC graphs to capture complementary information.
• To enhance the expressive power of GCNs by integrating dynamic topological variations.

Main Methods:

• Developed a dual-branch TSD-GCN architecture to model static and dynamic FC patterns.
• Static branch: adaptive topology learning on sFC using neural network layers.
• Dynamic branch: learning differential information across time steps to refine dynamic topology.

Main Results:

• TSD-GCN outperformed state-of-the-art methods on ADNI and ABIDE datasets.
• The dual-branch approach effectively integrated static and dynamic FC information.
• Discovered biologically meaningful discriminative FC patterns.

Conclusions:

• The TSD-GCN method offers a powerful approach for brain disorder classification by leveraging both static and dynamic FC topologies.
• Adaptive topology learning and integration of dynamic variations significantly enhance GCN performance.
• The findings highlight the importance of dynamic brain connectivity in disease recognition.