Research map / Graph learning / Knowledge and temporal graphs
Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis
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 WiKG, a dynamic directed graph representation for WSI classification that builds top-k neighbors from learned head/tail embeddings and aggregates them with a knowledge-aware attention mechanism.
Problem
Weakly supervised WSI classification with MIL ignores inter-patch interactions, and graph methods rely on fixed spatial topology that limits interactions between distant patches and treats graphs as undirected.
Approach
Replace spatial graph construction with dynamic top-k neighbor selection from learnable head and tail projections, form directed edge embeddings, and update head nodes via knowledge-graph-style attention over (head, relation, tail) triplets, followed by global readout.
Key results
- On TCGA-ESCA typing, WiKG reaches 95.23% AUC versus 93.79% for the best prior baseline (DTFD-MIL and HIPT).
- On TCGA-LUNG typing, WiKG achieves 90.78% AUC and 84.02% accuracy, outperforming DTFD-MIL (88.70% AUC).
- On the in-house FROZEN-LUNG external test set, WiKG reaches 87.06% accuracy, the best among compared methods, with lower per-class variance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WiKG instead of the strongest listed baseline | TCGA-ESCA typing · TCGA-ESCA typing | AUC: 93.980.83 -> 95.232.90 | Helped |
| Use WiKG instead of the strongest listed baseline | TCGA-KIDNEY typing · TCGA-KIDNEY typing | AUC: 99.550.08 -> 99.650.12 | Helped |
| Use WiKG instead of the strongest listed baseline | TCGA-LUNG typing · TCGA-LUNG typing | AUC: 89.251.87 -> 90.781.24 | Helped |
| Use WiKG instead of the strongest listed baseline | TCGA-ESCA staging · TCGA-ESCA staging | F1-score: 51.563.76 -> 55.804.39 | Helped |
| Use WiKG instead of the strongest listed baseline | TCGA-KIDNEY staging · TCGA-KIDNEY staging | AUC: 68.962.22 -> 69.711.40 | Helped |
| Use WiKG instead of the strongest listed baseline | TCGA-LUNG staging · TCGA-LUNG staging | AUC: 57.932.74 -> 60.341.37 | Helped |
| Construct directed edges using head-tail similarity rather than k-NN cosine similarity (ablation) | WiKG · TCGA-ESCA typing | AUC: 94.982.52 -> 95.232.90 | Helped |
| Construct directed edges using head-tail similarity rather than k-NN cosine similarity (ablation) | WiKG · TCGA-ESCA staging | AUC: 64.364.38 -> 69.964.28 | Helped |
| Construct directed edges using head-tail similarity rather than k-NN cosine similarity (ablation) | WiKG · TCGA-KIDNEY typing | AUC: 99.440.28 -> 99.650.12 | Helped |
| Construct directed edges using head-tail similarity rather than k-NN cosine similarity (ablation) | WiKG · TCGA-KIDNEY staging | AUC: 68.621.97 -> 69.711.40 | Helped |
| Replace knowledge-aware attention with GCN (ablation) | WiKG · TCGA-ESCA typing | AUC: 94.242.84 -> 95.232.90 | Helped |
| Replace knowledge-aware attention with GIN (ablation) | WiKG · TCGA-ESCA typing | AUC: 93.582.37 -> 95.232.90 | Helped |
Surprising findings
- Several MIL baselines show large gaps between AUC and accuracy on the external FROZEN-LUNG set, e.g., TransMIL has high AUC but low accuracy/F1.
Limitations
- Staging performance remains modest (around 60-70% accuracy), and the authors note staging adapts typing-oriented algorithms.
- Uses an ImageNet-pretrained ViT-Small encoder; domain-specific pathology encoders are only in supplementary material.
Experimental setup
Datasets: TCGA-ESCA TCGA-KIDNEY TCGA-LUNG FROZEN-LUNG (in-house) Camelyon16
Benchmarks: TCGA-ESCA TCGA-KIDNEY TCGA-LUNG FROZEN-LUNG
Metrics: accuracy AUC F1-score
Baselines: ABMIL CLAM-SB CLAM-MB DSMIL TransMIL DTFD-MIL HIPT GTP Patch-GCN GCN GIN GraphSAGE
Builds on: ABMIL knowledge graph attention network (KGAT) Patch-GCN CLAM
Closest papers by meaning
- Rethinking Transformer for Long Contextual Histopathology Whole Slide Image Analysis (NeurIPS 2024 · similarity 0.91)
- Sequential Attention-based Sampling for Histopathological Analysis (NeurIPS 2025 · similarity 0.88)
- Dynamic Entity-Masked Graph Diffusion Model for Histopathology Image Representation Learning (AAAI 2025 · similarity 0.88)
- Nonparametric Distribution Matching for Self-Supervised Whole-Slide Image Condensation (NeurIPS 2026 · similarity 0.88)
- HistoFS: Non-IID Histopathologic Whole Slide Image Classification via Federated Style Transfer with RoI-Preserving (CVPR 2025 · similarity 0.88)
- Unsupervised Foundation Model-Agnostic Slide-Level Representation Learning (CVPR 2025 · similarity 0.87)
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.