Research map / Graph learning / Knowledge and temporal graphs
Leveraging Tumor Heterogeneity: Heterogeneous Graph Representation Learning for Cancer Survival Prediction in Whole Slide 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.
Proposes ProtoSurv, a heterogeneous graph model for WSI survival prediction that uses tissue-category node labels from a fine-tuned UNI classifier and multi-prototype learning with prior-guided fusion, evaluated on five TCGA cancer cohorts.
Problem
Whole slide image survival prediction ignores intratumor tissue heterogeneity and the differing prognostic value of tissue types, while standard GCN/MIL methods assume homophily and lack pathological priors.
Approach
Builds a heterogeneous patch graph with tissue-category node labels and decouples it into a Structure View (multi-hop GCN) and a Histology View that learns multiple prototypes per prognostic tissue category (tumor, stroma, immune infiltration, necrosis, others), fused via prior-guided cross-attention and trained with Cox, compatibility, and orthogonality losses.
Key results
- ProtoSurv achieves C-index 0.774 on LGG, 0.720 on BRCA, 0.692 on COAD, outperforming the strongest baselines (e.g., 0.748 PANTHER and 0.746 HGSurvNet on LGG).
- On PAAD, PANTHER (0.673) slightly beats ProtoSurv (0.669), which the authors attribute to overfitting on the small 208-case dataset.
- Replacing the patch classifier with K-means (n=8) still outperforms most baselines, e.g., 0.656 on PAAD and 0.664 on COAD, showing robustness to classifier choice.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ProtoSurv instead of PANTHER | UNI-based survival prediction · PAAD | C-index: 0.673 -> 0.669 | Hurt |
| use ProtoSurv instead of HGSurvNet | UNI-based survival prediction · BRCA | C-index: 0.701 -> 0.720 | Helped |
| use ProtoSurv instead of PANTHER | UNI-based survival prediction · LGG | C-index: 0.748 -> 0.774 | Helped |
| use ProtoSurv instead of ABMIL | UNI-based survival prediction · LUAD | C-index: 0.653 -> 0.658 | Helped |
| use ProtoSurv instead of TransMIL | UNI-based survival prediction · COAD | C-index: 0.695 -> 0.692 | Hurt |
| add Histology View (ablation) | ProtoSurv · PAAD | C-index: 0.618 -> 0.669 | Helped |
| add Histology View (ablation) | ProtoSurv · BRCA | C-index: 0.647 -> 0.720 | Helped |
| add Histology View (ablation) | ProtoSurv · LGG | C-index: 0.713 -> 0.774 | Helped |
| add Histology View (ablation) | ProtoSurv · LUAD | C-index: 0.635 -> 0.658 | Helped |
| add Histology View (ablation) | ProtoSurv · COAD | C-index: 0.652 -> 0.692 | Helped |
| add Structure View with mean pooling (ablation) | ProtoSurv · BRCA | C-index: 0.657 -> 0.720 | Helped |
| add Structure View with mean pooling (ablation) | ProtoSurv · LGG | C-index: 0.706 -> 0.774 | Helped |
Surprising findings
- Simply clustering patches with K-means instead of a specialized tissue classifier still beats most comparison methods.
- Detailed 12-category tissue labels generally hurt performance compared with the coarse prior-based 5-category setting.
Limitations
- Requires patch-level tissue category labels from an additional classifier trained on annotated data, limiting applicability.
- Does not consistently beat the best baseline on PAAD, where small sample size may cause overfitting.
- Higher computational cost (627 GFLOPs, 39.1M parameters) than PatchGCN.
Experimental setup
Datasets: TCGA-BRCA TCGA-LGG TCGA-LUAD TCGA-COAD TCGA-PAAD CONCH/UNI tissue classification datasets
Benchmarks: TCGA-BRCA TCGA-LGG TCGA-LUAD TCGA-COAD TCGA-PAAD
Metrics: C-index Kaplan-Meier log-rank p-value
Baselines: WSISA ABMIL TransMIL DeepAttnMISL Patch-GCN DeepGraphConv HEAT HGSurvNet PANTHER
Builds on: Patch-GCN PANTHER HEAT UNI prototype-based heterogeneous graph learning (Dong et al.)
Closest papers by meaning
- Tumor Micro-environment Interactions Guided Graph Learning for Survival Analysis of Human Cancers from Whole-slide Pathological Images (CVPR 2024 · similarity 0.90)
- Multimodal Prototyping for cancer survival prediction (ICML 2024 · similarity 0.88)
- Modeling Dense Multimodal Interactions Between Biological Pathways and Histology for Survival Prediction (CVPR 2024 · similarity 0.87)
- Interpretable Vision-Language Survival Analysis with Ordinal Inductive Bias for Computational Pathology (ICLR 2025 · similarity 0.86)
- From Representation Space to Prognostic Insights: Whole Slide Image Generation with Hierarchical Diffusion Model for Survival Prediction (AAAI 2025 · similarity 0.86)
- Prototypical Information Bottlenecking and Disentangling for Multimodal Cancer Survival Prediction (ICLR 2024 · similarity 0.86)
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.