| Autor: |
Jiujiang Guo, Monan Wang, Xiaojing Guo |
| Jazyk: |
angličtina |
| Rok vydání: |
2024 |
| Předmět: |
|
| Zdroj: |
IEEE Access, Vol 12, Pp 162145-162156 (2024) |
| Druh dokumentu: |
article |
| ISSN: |
2169-3536 |
| DOI: |
10.1109/ACCESS.2024.3468439 |
| Popis: |
The diagnosis of neurodevelopmental disorders has been a challenge due to the heterogeneity in traits and high variations in observations. Functional Magnetic Resonance Imaging (fMRI) and its Functional Connectivity (FC) analysis have become instrumental in studying these disorders by accessing underlying abnormal neural function and communications. Recently, graph neural networks (GNNs) have shifted the analysis of brain networks by capturing the structural information contained within Regions of Interest (ROIs) interactions. Nonetheless, most existing studies are limited by oversimplifying complex brain connections and the nuanced relationships among brain regions. Additionally, the underlying interactions at multiple scales and semantics are often ignored. In this regard, we propose the Hierarchical Hyperedge Graph Transformer network (HH-GraphFormer), which builds hierarchical interactions of the brain and generates hyperedge graphs to detail relationships at various scales. Based on hyperedges formed from the doubly stochastic matrices of self-attention maps of the Transformer, multiple sub-graphs are constructed and used to depict diverse inter-ROI dependencies. Moreover, to aggregate the topological properties of multiple sub-graphs, we present a cluster-based aggregation module to gather sub-graph embeddings. Experimental results on the ABIDE and ADHD-200 datasets demonstrated the superior performance of HH-GraphFormer in diagnosing neurodevelopmental disorders and interpreting complex interactions in the brain. |
| Databáze: |
Directory of Open Access Journals |
| Externí odkaz: |
|
