Generative Hypergraph Neural Networks for Data Fusion | **Oral Presentation** | PRIME-MICCAI 2024
#GNN #HypergraphFusion #MICCAI2024
"DHSampling: Diversity-based Hyperedge Sampling in GNN Learning with Application to Medical Image Classification", Workshop on MLMI, MICCAI 2024, [Jiameng Liu, Furkan Pala, Islem Rekik, and Dinggang Shen]
👉 Paper link: https://link.springer.com/chapter/10....
👉 arXiv link: TO-UPDATE
👉 Code link: https://github.com/basiralab/ Gen-HNN
A connectional brain template (CBT) is a fingerprint graph-based representation of a population of brain networks, serving as an ‘average’ connectome. CBTs are essential for creating representative maps of brain connectivity in both typical and atypical populations, facilitating the identification of deviations from healthy brain structures. However, traditional methods for generating CBTs often rely on linear averaging and pairwise relationships, which fail to capture the complex, high-order interactions within brain networks, particularly in multi-view brain networks where the brain is encoded in a set of connectivity matrices (i.e., tensor). To address these limitations, we propose a novel Generative Hypergraph Neural Network (Gen-HNN) for learning hyper connectional brain templates (HCBTs). Gen-HNN leverages hypergraphs to capture higher-order relationships, utilizing hyperedge convolution operations based on the hypergraph Laplacian to process and integrate multi-view brain data into a cohesive HCBT. Our model overcomes the limitations of existing methods by effectively handling non-linear pat- terns and preserving the topological properties of brain networks. We conducted extensive experiments, demonstrating that Gen-HNN significantly outperforms state-of-the-art methods in terms of both representativeness and discriminative power.