Abstract
Survival prediction is a complex ordinal regression task that aims to predict the survival coefficient ranking among a cohort of patients, typically achieved by analyzing patients' whole slide images (WSIs). Existing deep learning approaches mainly adopt multiple instance learning or graph neural networks under weak supervision. However, most are unable to uncover the diverse interactions between different types of biological entities (e.g., cell clusters and tissue blocks) across multiple scales, while such interactions are crucial for patient survival prediction. In light of this, we propose a novel multi-scale heterogeneity-aware hypergraph representation framework (H2GT). Specifically, our framework first constructs a multi-scale heterogeneity-aware hypergraph and assigns each node its biological entity type. It then mines diverse interactions between nodes on the graph structure to obtain a global representation. Experimental results demonstrate that our method outperforms state-of-the-art approaches on three benchmark datasets.
Introduction
Predicting patient prognosis from Whole Slide Images (WSIs) is a cornerstone of modern computational pathology. WSIs are massive digital files capturing tissue biopsies at sub-micron resolution, often exceeding 40,000 x 40,000 pixels. The primary challenge lies in extracting meaningful prognostic signals from these gargantuan images.
Current AI approaches typically decompose WSIs into small patches, treating them either as a "bag of instances" (Multiple Instance Learning) or as nodes in a graph (Graph Neural Networks). However, these methods often treat all patches as homogeneous, ignoring the complex biological ecosystem of a tumor. In reality, the interaction between a tumor cell and an immune cell is fundamentally different from the interaction between two tumor cells. Existing models struggle to capture these diverse, higher-order interactions across different scales, which are critical indicators of patient survival. We address this by introducing a framework that explicitly models biological heterogeneity through hypergraphs.
Methodology
The proposed Heterogeneous HyperGraph Transformer (H2GT) operates through a sophisticated three-stage pipeline designed to preserve biological context across scales.
1. Pairwise Heterogeneous Graph Construction
The WSI is first segmented into patches, which are classified into six distinct biological categories (e.g., Neoplastic, Inflammatory, Stroma, Necrosis). A heterogeneous graph is constructed where nodes represent patches and edges represent the Pearson correlation between their extracted features, conditioned on their tissue types.
2. Heterogeneity-Aware Hypergraph
To capture relationships beyond simple pairs, we utilize hypergraphs where a single hyperedge can connect multiple nodes. We specifically define "Mix Nodes"—patches containing multiple tissue types—and "Boundary Hyperedges" to model the tumor-stroma invasion zone. This allows the model to explicitly focus on the invasive front of the tumor, a known pathological marker for poor prognosis.
3. H2GT Transformer
The core of the architecture is the Heterogeneous Multi-Head Attention (HMHA) layer. Unlike standard transformers, HMHA uses unique projection matrices for different node and hyperedge types. This ensures that the signal from an inflammatory region is processed with different weights than a necrotic region, preserving the semantic heterogeneity of the tissue throughout the learning process.
Experimental Results
We evaluated H2GT on three large-scale benchmark datasets from The Cancer Genome Atlas (TCGA): BRCA (Breast Invasive Carcinoma), GBMLGG (Glioblastoma & Lower Grade Glioma), and BLCA (Bladder Urothelial Carcinoma). The study processed approximately 25 million image patches across nearly 2,000 WSIs.
Performance and Stratification
H2GT achieved state-of-the-art (SOTA) performance, reaching an average C-index of 66.08%, surpassing previous methods like HEAT (64.94%) and PatchGCN (64.05%). Beyond ranking, the model demonstrated exceptional ability in patient stratification. Kaplan-Meier analysis showed that H2GT could successfully separate patients into high-risk and low-risk groups with high statistical significance (p < 0.05) across all tested cancer types.
Conclusion
H2GT represents a significant advancement in WSI analysis by shifting from homogeneous patch processing to heterogeneity-aware hypergraph modeling. By explicitly representing the diverse biological entities and their higher-order interactions, the model aligns more closely with clinical pathology principles. While computationally intensive, the framework's ability to identify critical prognostic regions like the invasive front marks a major step toward more interpretable and accurate AI-driven cancer prognosis.
Summary updated