Graph Neural Network (GNN)

The core operation where nodes iteratively aggregate feature information from their neighbors to update their own representations. In digital pathology, a tumor cell node receives signals from adjacent immune and stromal cells, learning its contextual identity within the tissue architecture. This process typically runs for 2-5 message-passing layers, balancing local neighborhood information with computational tractability. The aggregation function—whether mean, sum, or attention-weighted—determines how influence propagates through the cellular graph.

Raw cell centroid coordinates are converted into a graph by defining edges based on spatial proximity. Common strategies include:

• K-Nearest Neighbors (KNN): Connect each cell to its K closest neighbors, preserving local density variations
• Radius-based: Connect all cells within a fixed distance threshold, capturing biological interaction ranges
• Delaunay triangulation: Creates a planar graph that naturally respects tissue boundaries The choice of construction method directly impacts which spatial patterns the GNN can detect.

Each cell node is initialized with a feature vector encoding its intrinsic properties before message passing begins. Features typically include:

• Morphological measurements: area, perimeter, eccentricity, solidity
• Intensity statistics: mean H&E channel values, texture descriptors
• Biomarker expression: quantified IHC or multiplex IF staining intensities
• Cell type predictions: one-hot encoded classifications from upstream classifiers These features provide the initial signal that the GNN refines through spatial context.

After message passing, node embeddings are aggregated into a single graph-level representation for slide-level prediction tasks. Common readout functions include:

• Global mean pooling: averages all node features, simple but loses spatial heterogeneity
• Global max pooling: captures the strongest signals across the graph
• Attention-based pooling: learns to weight diagnostically relevant regions
• Hierarchical pooling: progressively coarsens the graph, preserving multi-scale structure This readout vector feeds into a classifier for tasks like survival prediction or treatment response.

Graph Attention Networks (GATs) extend standard GNNs by learning edge-specific importance weights during message passing. Rather than treating all neighbors equally, the model computes attention coefficients that amplify signals from biologically relevant interactions—such as a CD8+ T cell directly contacting a tumor cell—while suppressing noise from distant or irrelevant cells. This mechanism provides inherent interpretability, as attention weights can be visualized to reveal which spatial relationships drive predictions.

The constructed graph directly encodes the tumor microenvironment topology that pathologists assess qualitatively. Edges between tumor cells and tumor-infiltrating lymphocytes (TILs) capture immune infiltration patterns. Clusters of fibroblasts connected to collagen-rich regions represent desmoplastic stroma. By learning on these explicit relational structures, GNNs quantify spatial biomarkers—such as immune exclusion versus infiltration—that correlate with immunotherapy response and patient prognosis.

Immune cells whose spatial distribution and clustering patterns are naturally modeled by GNNs. Rather than simply counting TIL density, graph-based approaches capture the topological relationships between lymphocytes and tumor cells—quantifying immune exclusion, infiltration, and tertiary lymphoid structures as predictive biomarkers for immunotherapy response.

• Models immune cell spatial architecture
• Quantifies immune exclusion vs. infiltration
• Predicts checkpoint inhibitor response

A weakly-supervised paradigm often combined with GNNs for slide-level prediction. While standard MIL treats patches as independent instances, graph-based MIL constructs a graph where patches are nodes connected by spatial adjacency, allowing the model to reason about tissue architecture rather than isolated regions. This hybrid approach improves cancer subtyping and grading accuracy.

• Patches as nodes with spatial edges
• Captures long-range tissue dependencies
• Outperforms standard attention MIL

A visualization technique that reveals which graph nodes and edges most influenced a GNN's prediction. In pathology, attention heatmaps highlight diagnostically relevant tissue regions and the spatial relationships between them, providing interpretability for regulatory approval. Graph attention weights can identify specific cell-cell interactions driving the model's decision.

• Highlights influential cell neighborhoods
• Reveals key cell-cell interactions
• Supports FDA submission interpretability

The high-throughput extraction of quantitative features from digital pathology, where graph-based pathomics captures topological features invisible to conventional texture analysis. GNNs compute graph-level descriptors—such as clustering coefficients, centrality measures, and community structure—that characterize tumor heterogeneity and stromal architecture as prognostic biomarkers.

• Extracts topological tissue features
• Quantifies tumor-stroma organization
• Complements traditional radiomics

A complementary architecture that applies self-attention to image patches, often used alongside GNNs in hybrid models. While ViTs capture long-range visual dependencies within a tissue region, GNNs explicitly model the relational structure between detected objects—cells, glands, or tissue compartments—making the combination powerful for holistic tissue understanding.

• ViT for patch-level feature extraction
• GNN for inter-object relationship modeling
• Hybrid architectures achieve state-of-the-art