TopoMamSurv: Graph Mamba Framework Boosts Pathology Survival Analysis
Linear-complexity Mamba model with topology-aware ordering outperforms Transformers on 5 cancer datasets.
Survival analysis from Whole Slide Images (WSIs) is critical for cancer prognosis, but existing models struggle with scale. Transformers capture long-range dependencies but suffer O(N²) complexity, making them impractical for large WSI graphs. Mamba models offer linear complexity but are highly sensitive to input order—traditional node sorting (by degree or subgraph size) ignores topological connectivity, limiting performance.
To solve this, the authors propose TopoMamSurv, a novel Graph Mamba framework. Its core innovation is a topology-aware ordering (TAO) strategy that arranges nodes to maximize similarity along the sequence, improving Mamba's sequential modeling. A bidirectional Mamba module then captures spatial context from both directions, while a Graph Convolutional Network (GCN) handles local feature aggregation. This forms a hierarchical architecture: local aggregation → global capture.
Experiments on five TCGA datasets show TopoMamSurv achieves comprehensive performance advantages over both Transformer and baseline Mamba models. Visualization confirms TAO produces higher node similarity. The framework effectively reconciles long-range dependency modeling, computational efficiency, and spatial structure utilization—a significant step for scalable AI in digital pathology.
- Topology-Aware Ordering (TAO) arranges graph nodes to maximize sequence similarity, improving Mamba's state-space modeling.
- Bidirectional Mamba module enables dual-direction spatial context capture, combining with GCN for hierarchical feature learning.
- Validated on five TCGA cancer datasets, achieving linear O(N) complexity vs. Transformers' O(N²).
Why It Matters
Faster, more accurate cancer prognosis from massive pathology images, enabling scalable AI for clinical decision support.