Research map / Graph learning / Knowledge and temporal graphs
Tumor Micro-environment Interactions Guided Graph Learning for Survival Analysis of Human Cancers from Whole-slide Pathological Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds spatial graphs of tumor, lymphocyte and stromal patches from WSIs, learns TME-neighborhood-organization-aware node embeddings, and applies a gated graph attention network with a Cox model for survival prediction on TCGA cohorts.
Problem
Survival prediction from whole-slide pathology images that existing methods make from randomly or tumor-selected patches, which cannot capture interactions between tumor and its micro-environment (lymphocytes, stromal fibrosis).
Approach
Node embeddings are trained so that their TME neighborhood organization similarity (k-hop composition of tumor/lymphocyte/stromal neighbors, clustered into groups, matched via KL divergence) is preserved, and a gated graph attention network models same-type (GGC) and cross-type (GAT) interactions before Cox-based survival prediction.
Key results
- TMEGL achieves CI of 0.719 (BRCA), 0.697 (KIRC) and 0.695 (LUSC), versus best baseline HGSurvNet at 0.671, 0.659 and 0.667.
- AUC of 0.750, 0.716 and 0.736 on BRCA, KIRC and LUSC, outperforming HGSurvNet (0.681, 0.681, 0.701).
- Ablation: removing the graph embedding step (TMEGL-NE) drops CI to 0.609 on BRCA, versus 0.719 for full TMEGL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TMEGL rather than HGSurvNet | WSI survival prediction · BRCA | CI: 0.671(0.04) -> 0.719(0.03) | Helped |
| Use TMEGL rather than HGSurvNet | WSI survival prediction · KIRC | CI: 0.659(0.04) -> 0.697(0.03) | Helped |
| Use TMEGL rather than Co-Pilot | WSI survival prediction · LUSC | CI: 0.674(0.03) -> 0.695(0.04) | Helped |
| Add TME neighborhood organization guided graph embedding instead of omitting it or using DeepWalk (ablation) | TMEGL · BRCA, KIRC, LUSC | CI: TMEGL: 0.719, 0.697, 0.695; TMEGL-NE: 0.609, 0.629, 0.618; TMEGL-DW: 0.658, 0.631, 0.621 | Helped |
| Distinguish same-type from different-type edges in the graph attention network (ablation) | TMEGL · BRCA, KIRC, LUSC | CI: TMEGL: 0.719, 0.697, 0.695; TMEGL-GAT: 0.662, 0.630, 0.633 | Helped |
| Use neighborhood organization across 1-, 2-, and 3-hop ranges rather than only 1- or 2-hop (ablation) | TMEGL · BRCA, KIRC, LUSC | CI: TMEGL: 0.719, 0.697, 0.695; TMEGL-N1: 0.644, 0.641, 0.634; TMEGL-N2: 0.671, 0.663, 0.657 | Helped |
| Aggregate each node type into B = 4 groups rather than B = 2 (ablation) | TMEGL · BRCA, KIRC, LUSC | CI: TMEGL: 0.719, 0.697, 0.695; TMEGL-G2: 0.643, 0.633, 0.641 | Helped |
| Aggregate each node type into B = 4 groups rather than B = 5 (ablation) | TMEGL · BRCA, KIRC, LUSC | CI: TMEGL: 0.719, 0.697, 0.695; TMEGL-G5: 0.694, 0.680, 0.681 | Helped |
| Combine intra-group and inter-group similarity rather than using intra-group similarity only (ablation) | TMEGL · BRCA, KIRC, LUSC | CI: TMEGL: 0.719, 0.697, 0.695; TMEGL-Intra: 0.676, 0.636, 0.641 | Helped |
| Combine intra-group and inter-group similarity rather than using inter-group similarity only (ablation) | TMEGL · BRCA, KIRC, LUSC | CI: TMEGL: 0.719, 0.697, 0.695; TMEGL-Inter: 0.668, 0.623, 0.646 | Helped |
Limitations
- Relies on a U-Net++ tissue segmentation model with dice 0.858 for patch-type labels, so segmentation errors propagate to graph construction.
- Only three TME component types (tumor, lymphocytes, stromal) and 300 patches per type are used, limiting coverage of the microenvironment.
- Patch selection still depends on a pretrained segmentation model rather than fully annotation-free pipelines.
Experimental setup
Datasets: TCGA-BRCA TCGA-KIRC TCGA-LUSC annotated tissue segmentation dataset from [1]
Benchmarks: TCGA-BRCA TCGA-KIRC TCGA-LUSC
Metrics: concordance index AUC log-rank p-value
Baselines: WSISA DeepAttnMISL MCAT DeepGraphSurv Patch-GCN CluSiam Co-Pilot HGSurvNet
Builds on: Gated Graph Neural Networks SAGPool DeepWalk DeepGraphSurv
Closest papers by meaning
- Leveraging Tumor Heterogeneity: Heterogeneous Graph Representation Learning for Cancer Survival Prediction in Whole Slide Images (NeurIPS 2024 · similarity 0.90)
- Modeling Dense Multimodal Interactions Between Biological Pathways and Histology for Survival Prediction (CVPR 2024 · similarity 0.87)
- Multi-modal Topology-embedded Graph Learning for Spatially Resolved Genes Prediction from Pathology Images with Prior Gene Similarity Information (CVPR 2025 · similarity 0.86)
- MERGE: Multi-faceted Hierarchical Graph-based GNN for Gene Expression Prediction from Whole Slide Histopathology Images (CVPR 2025 · similarity 0.86)
- Multimodal Prototyping for cancer survival prediction (ICML 2024 · similarity 0.86)
- Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.