Cross-Group Brain Network Topology Propagation Model with Graph Attention Mechanism

Most brain diseases are characterized by widespread structural and functional brain abnormalities. Exploring the complex changes in network topology caused by brain diseases has unique advantages in revealing neuropathological disease characteristics. Current brain network analysis methods mainly fo...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:2023 7th Asian Conference on Artificial Intelligence Technology (ACAIT) S. 421 - 426
Hauptverfasser: Yu, Jing, Li, Shengrong, Ma, Kai, Wan, Peng, Sun, Liang, Zhu, Qi
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 10.11.2023
Schlagworte:
Adaptation models
Biological system modeling
Degradation
Epilepsy
Functional connectivity
Graph autoencoder
Group differences
Network topology
Predictive models
Structural connectivity
Weight measurement
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Most brain diseases are characterized by widespread structural and functional brain abnormalities. Exploring the complex changes in network topology caused by brain diseases has unique advantages in revealing neuropathological disease characteristics. Current brain network analysis methods mainly focus on the low-order association of paired brain regions, ignoring the complicated high-order information among multiple brain regions. In this work, we propose a cross-group brain network topology propagation model with a graph attention mechanism, which constructs a high-order cross-group brain network with group differences. Firstly, we extract the functional brain network topology of healthy subjects and patients and adaptively evaluate the centrality distribution of brain regions by the PageRank algorithm. Secondly, based on the optimal transport, the node centrality propagation model of brain regions from two groups is established. The constructed transmission matrix captures higher-order topological features with group differences among multiple brain regions caused by brain diseases. Then, a Graph Autoencoder model (GAE) based on the Graph Attention Network (GAT) is designed to embed the topological structure of functional connectivity fusing cross-group brain network topology. Structural connectivity is reconstructed after decoding the encoded feature. Finally, the embedding feature is utilized for an individual classification task. Experiments on an epilepsy dataset show that our method is not only capable of identifying biomarkers, but also surpasses several state-of-the-art methods for diagnosing brain diseases.
DOI:10.1109/ACAIT60137.2023.10528555