Title: TopoMamSurv: A Topology-Aware Ordered Graph Mamba Framework for Survival Analysis

Abstract:

In the field of computational pathology, survival analysis using Whole Slide Images (WSIs) plays a pivotal role in assessing patient prognosis. However, this process is hindered by several technical hurdles. While Transformers can capture long-range dependencies via self-attention, their $O(N^2)$ time complexity creates a significant computational bottleneck when processing the large-scale graph structures inherent in WSIs. The Mamba model offers a solution to this bottleneck by operating with linear complexity. Nevertheless, Mamba is highly sensitive to the sequence of input data. Traditional sorting methods used in Graph Mamba—such as ordering by node degree or subgraph size—often overlook the topological connectivity of the graph, thereby limiting the effectiveness of Mamba’s sequential modeling capabilities. Additionally, Mamba’s standard unidirectional design fails to exploit the bidirectional spatial structures found in images.

To overcome these limitations, this study introduces TopoMamSurv, a novel Graph Mamba survival analysis framework driven by topology-aware ordering (TAO). This approach is specifically designed to accommodate Mamba’s sequential sensitivity. Visualization experiments demonstrate that nodes identified through the TAO strategy possess significantly higher similarity. Furthermore, we developed a bidirectional Mamba module and incorporated a Graph Convolutional Network (GCN) to enable bidirectional spatial context modeling. This integration establishes a hierarchical feature learning architecture characterized by "local aggregation" and "global capture." Through the systematic implementation of TAO, bidirectional semantic modeling, and hierarchical feature fusion, the framework successfully balances the competing demands of long-range dependency modeling, computational efficiency, and spatial structure utilization in WSI analysis. The comprehensive performance superiority of this framework has been validated across five TCGA datasets.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC