Graph Neural Network (GNN)

The foundational step where a gigapixel Whole Slide Image (WSI) is transformed into a graph structure.

• Nodes: Represent individual tissue patches or superpixels, each encoded with a Feature Embedding extracted by a pre-trained encoder (e.g., a ViT).
• Edges: Connect nodes based on spatial proximity (e.g., k-nearest neighbors or a distance threshold), explicitly encoding the tissue topology.
• This process converts an unstructured pixel grid into a structured relational data format suitable for graph-based learning.

The core computational mechanism of a GNN, where nodes iteratively update their representations by communicating with neighbors.

• Message Function: A learned function computes a 'message' from a neighboring node, often conditioned on both node features and edge attributes.
• Aggregation Function: A permutation-invariant operation (e.g., sum, mean, or max) combines messages from all neighbors.
• Update Function: The node's current state is updated using the aggregated message, allowing contextual information to propagate across the graph, capturing long-range tissue interactions.

GNNs are uniquely suited to capture the complex ecology of the Tumor Microenvironment (TME).

• Tumor-Stroma Ratio: The model can learn the significance of the interface between malignant epithelium and surrounding stroma by analyzing edge connections between these distinct tissue classes.
• Tumor-Infiltrating Lymphocytes (TILs): The spatial clustering and dispersion of immune cells relative to the tumor boundary are naturally encoded in the graph's topology, providing a powerful prognostic signal.
• This explicit architectural modeling goes beyond simply counting cell types to understanding their functional organization.

After message passing, the entire graph must be summarized into a single vector for slide-level tasks like Slide-Level Classification.

• Global Pooling: A permutation-invariant readout function (e.g., global mean or max pooling) aggregates all node features into a single graph-level embedding.
• Hierarchical Pooling: Methods like DiffPool or Top-K pooling learn to coarsen the graph hierarchically, preserving important structural information while reducing to a final representation.
• This final embedding is passed to a standard classifier to predict the diagnostic label, such as cancer subtype or the presence of Microsatellite Instability (MSI).

GNNs can be seamlessly integrated with Multiple Instance Learning (MIL) frameworks for weakly supervised tasks.

• In a standard MIL setup, patches are treated as independent instances. A GNN augments this by making instances interdependent through their spatial connections.
• Attention-based MIL with GNNs: A GNN first refines patch embeddings with spatial context. An Attention Mechanism then weights these context-aware embeddings, allowing the model to focus on diagnostically relevant, spatially-informed regions.
• This hybrid approach, exemplified by frameworks like CLAM adapted with graph layers, often outperforms methods that ignore tissue architecture.

GNNs offer unique avenues for Algorithmic Explainability by identifying the critical structural motifs driving a prediction.

• GNNExplainer: A model-agnostic method that identifies a compact subgraph and a small subset of node features most influential for a specific prediction.
• Concept-Based Explanations: Instead of highlighting individual pixels (like Grad-CAM), GNN explanations can reveal higher-level concepts like 'a dense cluster of tumor cells adjacent to necrotic tissue,' which aligns more closely with pathologist reasoning.
• This structural interpretability is vital for validating model behavior and building clinical trust in its architectural assessments.

The process of converting a whole slide image into a graph where nodes represent individual tissue patches or cells, and edges encode spatial proximity. Delaunay triangulation or k-nearest neighbors algorithms connect neighboring patches, explicitly preserving the tissue architecture that convolutional approaches often ignore. Edge features can encode distance, orientation, or morphological similarity between connected regions.

The core computational mechanism of GNNs where nodes iteratively aggregate and transform information from their neighbors. In each layer, a node's representation is updated by combining its own features with messages from adjacent nodes using differentiable functions. This allows contextual information about the local cellular neighborhood to propagate across the graph, enabling the model to learn complex spatial interaction patterns like immune cell infiltration boundaries.

The operation that aggregates all node embeddings into a single graph-level representation for slide-level classification. Common strategies include:

• Global mean pooling: Averages all node features
• Global max pooling: Selects the most activated features
• Attention-based pooling: Learns to weight diagnostically relevant nodes
• Hierarchical pooling: Progressively coarsens the graph structure The choice of readout function significantly impacts the model's ability to capture global tissue architecture patterns.

Two primary granularities for graph construction in computational pathology:

Cell-Graphs: Nodes represent individually segmented cells (nuclei). Captures fine-grained cell-to-cell interactions but requires accurate nuclear segmentation and is computationally intensive with millions of nodes per WSI.

Patch-Graphs: Nodes represent larger tissue tiles (e.g., 256×256 pixels at 20× magnification). More computationally tractable and compatible with standard feature embeddings from pre-trained CNNs or vision transformers, but loses sub-cellular resolution.