A Hierarchical Hypergraph Attention Network for Survival Analysis from Pathological Images

The histopathology whole slide image (WSI)-based survival analysis has attracted increasing attention in recent years. However, existing computational pathology methods still encounter two major obstacles. First, complex data associations between histologic entities in WSI are difficult to mine for discovering heterogeneity. Second, global features in high-resolution WSI lack integration to perform comprehensive analysis. To address these challenges, we propose a novel deep neural network, termed Hierarchical Hypergraph Attention Network (HHGAT). Specifically, we depict the complex correlations among histological entities from both the morphological and topological perspectives, in which processes hypergraph neural network is incorporated for empowering the model's feature engineering capability. Additionally, HHGAT considers a hierarchical architecture and an attention-based multilevel aggregator to generate more discriminative global representations. Extensive experiments on three benchmark datasets demonstrate that HHGAT outperforms existing state-of-the-art methods.