Hierarchical Multi-Scale Graph Learning with Knowledge-Guided Attention for Whole-Slide Image Survival Analysis

A research team led by Bin Xu and Yufei Zhou has introduced HMKGN (Hierarchical Multi-scale Knowledge-aware Graph Network), a breakthrough AI architecture for cancer survival prediction from whole-slide pathology images. Unlike conventional attention-based or graph-based multiple instance learning (MIL) approaches that either ignore spatial organization or rely on static handcrafted graphs, HMKGN enforces hierarchical structure with spatial locality constraints. The system creates local cellular-level dynamic graphs that aggregate spatially proximate patches within each region of interest, then builds a global slide-level dynamic graph that integrates ROI-level features into comprehensive WSI-level representations. This approach allows the model to capture both fine-grained cellular patterns and broader tissue architecture relationships that are crucial for accurate prognosis.

The technical innovation lies in HMKGN's multi-scale integration at the ROI level, which combines coarse contextual features from broader views with fine-grained structural representations from local patch-graph aggregation. When evaluated on four TCGA cancer cohorts—kidney renal clear cell carcinoma (KIRC, N=513), lower-grade glioma (LGG, N=487), pancreatic adenocarcinoma (PAAD, N=138), and stomach adenocarcinoma (STAD, N=370)—the model consistently outperformed existing MIL-based approaches. It achieved a 10.85% improvement in concordance indices and produced statistically significant stratification of patient survival risk (log-rank p < 0.05). The research, accepted for ISBI 2026, represents a significant advancement in computational pathology that could enable more precise, data-driven cancer prognosis and treatment personalization.