HMHGT: Hierarchical Multi-Scale Hypergraph Representation for Histopathology Whole Slide Images

Cancer prognosis is a challenging task in computational pathology, requiring comprehensive representation of tissue features with context awareness to more accurately infer patient survival. However, existing deep learning methods, such as multi-instance learning and graph neural networks under weak supervision, often fail to uncover the complex interactions between biological entities across multiple scales (e.g., cell groups and tissue blocks). These methods typically either ignore spatial relationships entirely or model them through oversimplified pairwise connections, inadequately capturing the multi-entity interactions fundamental to understanding the tumor microenvironment. To address this issue, we propose a novel hierarchical multi-scale hypergraph transformer model (HMHGT). This framework constructs horizontal hypergraphs at two scales (20× for cellular details and 5× for tissue architecture) and establishes vertical connections to model higher-order interactions between different biological entities in the tumor microenvironment. Additionally, the attention-gated vertical fusion module aggregates instance-level histological features from multiple scales and effectively suppresses environmental noise through adaptive weighting. Experimental results on the Cancer Genome Atlas (TCGA) dataset demonstrate that HMHGT outperforms existing mainstream weakly supervised methods, achieving particularly strong performance on datasets with high tissue heterogeneity.