Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis

1. Introduction

• Digital scanning converts physical slides into high-resolution WSIs and has increased demand for intelligent diagnostic tools.The passage links scanning with improved pathologist work efficiency.Source: §1 · p. 1
• Embedding-based MIL aggregates patch information but can overlook interconnections that may inform tumor staging and clinical prognosis.The cited examples concern interactions between tumor and inflammatory cells.Source: §1 · pp. 1–2
• Existing graph methods rely on spatial topology, which can limit information exchange between distant entities and overlook directed contributions between adjacent patches.The introduction also notes that GNN over-parameterization may cause overfitting and reduce generalization.Source: §1 · p. 2
• WiKG models patch interactions with dynamic directed graphs and knowledge-aware attention, and the authors report superior performance across three TCGA datasets and in-house tests.Head and tail embeddings define directed relationships; attention aggregates neighbor information.Source: §1 · p. 2

2.1. Application of MIL in Histopathology WSIs

• Embedding-based MIL compresses patch embeddings and aggregates them for slide-level prediction, but the reviewed methods still lack informative interpatch interactions.ABMIL [19] uses attention-weighted instance aggregation; later work adds clustering constraints, dual-stream modeling, transformers, pseudo-bags, or feature distillation.Source: §2.1 · p. 2

2.2. Graph Representation in Digital Pathology

• Earlier graph work also represents tissue or cells as graphs, including spatial cell edges and combined tissue-cell representations for classification.Examples include HACT-Net [31] and SHGNN.Source: §2.2 · p. 2
• Figure 2 outlines the proposed WSI-analysis framework, from patch feature extraction through dynamic edge construction and graph learning to slide prediction.Dynamic edges are based on head and tail embeddings.Source: Figure 2 · p. 3
• Graph-based WSI methods model patch topology and interactions, but most use fixed spatial edges and do not fully account for directional contextual relations.The cited approaches include hierarchical patch aggregation, spatial representations with vision transformers, multiscale heterogeneous graphs, and multi-task graph-transformer models.Source: §2.2 · pp. 2–3

3.1. Dynamic Graph Construction

• For each patch, WiKG selects the top-k patches by head–tail similarity as neighbors and uses head and tail embeddings to construct directed edge embeddings.The resulting graph represents nodes, edges, head and tail embeddings, and directed edge embeddings.Source: §3.1 · p. 3

3.2. Knowledge-aware Attention Mechanism

• WiKG weights each neighbor’s tail information using a score derived from the head, relation, and tail, then aggregates it to update the head representation.The updated head fuses aggregated neighbor information with the original head through a dual-interaction mechanism.Source: §3.2 · p. 4
• A global pooling readout produces graph-level embeddings, which a softmax function converts into WSI class probabilities.The stated training objective is cross-entropy loss.Source: §3.2 · p. 4

4. Experiments

• The experiments compare WiKG with state-of-the-art WSI methods and include ablations of edge construction, knowledge-aware attention, neighbor count, convergence, and efficiency.Additional in-house testing evaluates generalization.Source: §4 · p. 4
• Table 1 reports cancer typing results for various methods across TCGA-ESCA, TCGA-KIDNEY, and TCGA-LUNG.Source: Table 1 · p. 5

4.1. Datasets

• The public cohorts are TCGA-ESCA (375 cases), TCGA-KIDNEY (1,233), and TCGA-LUNG (2,121), with tumor typing and staging tasks for each.Staging cases are categorized into stages I–IV according to TNM standards.Source: §4.1 · p. 5
• TCGA results use four-fold cross-validation, with AUC, Accuracy, and weighted F1-score reported; the in-house test includes 170 frozen-section cases.The in-house cases are from the First Affiliated Hospital of Sun Yat-sen University and include lung squamous cell carcinomas and adenocarcinomas.Source: §4.1 · p. 5

• Each patch is encoded as a 384-dimensional feature and projected to 512 dimensions; the setup uses a pretrained ImageNet ViT Small encoder.The encoder is selected for single-GPU use; pathology-pretrained encoder comparisons are in the supplementary material.Source: §4.2 · p. 5
• Training uses Topk k=6, batch size 1, 100 epochs, Adam, learning rate 10^-4, weight decay 10^-5, and dropout 0.3.Baselines use the same parameters and settings.Source: §4.2 · p. 5
• Table 3 reports cancer staging results for various methods on TCGA-ESCA, TCGA-KIDNEY, and TCGA-LUNG.Source: Table 3 · p. 6

4.3. Comparison with Other Existing Works

• WiKG demonstrates superior performance to MIL and graph representation methods on tumor typing and staging across three datasets.For staging, its improvement in F1-score is more significant than in Accuracy.Source: §4.3 · p. 6
• Current staging algorithms are generally adaptations of typing algorithms, and the authors identify staging-specific deep learning as future work.Source: §4.3 · p. 6

4.4. Effectiveness of Proposed Edge Construction

• WiKG's head-tail edge construction outperforms k-NN distance and cosine variants by nearly one to three percentage points in AUC across two tasks and all three datasets.The comparison changes edge construction while keeping the other structures unchanged.Source: §4.4 · p. 6, Table 4 · p. 6

• WiKG achieves superior AUC to GCN, GIN, SAGE, and GAT across three WSI typing tasks.Source: §4.5 · p. 7, Table 5 · p. 7
• Nearly 2%: WiKG improves AUC in esophageal-cancer analysis compared with the alternative GNN architectures.Source: §4.5 · p. 7, Table 5 · p. 7
• GAT also aggregates node features through attention scores, but performs worse than WiKG in the reported comparison.The authors suggest WiKG may better capture effective neighbor information.Source: §4.5 · p. 7

4.6. Effect of Numbers of Neighbor Nodes

• Across typing and staging on three TCGA datasets, the number of neighbor nodes has little impact on WiKG performance.Source: §4.6 · p. 7
• TCGA-ESCA staging Accuracy fluctuates with neighbor count; the authors consider this reasonable given uneven stage counts and possibly related to esophageal cancer's morphological diversity and complexity.Source: §4.6 · p. 7

4.7. Convergence and Time Efficiency

• WiKG converges quickly and shows no overfitting compared with Patch-GCN and GTP in TCGA-ESCA staging.Source: Figure 5 · p. 7, §4.7 · p. 7
• WiKG has the fastest single-epoch training speed and the smallest number of parameters in the comparison.Source: §4.7 · p. 7

Visualization

• In Camelyon16, metastatic-cancer patches tend to select lymphocyte-rich patches as neighbors, with adipose and fibrous tissue as supplements; the highest-attention patch contains more lymphomas.Source: §4.8 · pp. 7–8, Figure 6 · p. 8

Further Experiment

• On FROZEN-LUNG, WiKG has superior overall performance, recognizes both lung subtypes well, and has lower per-category Accuracy standard deviations than other methods.Source: §4.8 · p. 8, Table 6 · p. 8, Figure 7 · p. 8
• TransMIL has unusually high AUC despite relatively low Accuracy and F1-score; its squamous-cell carcinoma Accuracy is nearly 44 percent higher than its adenocarcinoma Accuracy, possibly reflecting training-set imbalance.Source: §4.8 · p. 8
• Other compared methods generally show large category-level Accuracy differences and high category-specific standard deviations, issues WiKG significantly mitigates.Source: §4.8 · p. 8

5. Conclusion and Future Work

• WiKG uses head-tail interactions and knowledge-aware attention to aggregate neighbor information, and experiments on three public benchmarks and in-house test sets demonstrate effectiveness and better generalization.Source: §5 · p. 8
• Future work will focus on WiKG's interpretability and graph-pooling effects, and further study generalization in WSI analysis tasks.Source: §5 · p. 8