Research on Tissue Microenvironment Modeling Using Graph Neural Networks for Lymph Node Metastasis Prediction in Colorectal Cancer

The bachelor's thesis by Enric Ferrera González from the Autonomous University of Barcelona explores the use of Graph Neural Networks (GNNs) to model the spatial tissue microenvironment of pT1-stage colorectal cancer for predicting lymph node metastasis risk. This article analyzes its graph construction strategies, GNN architecture design, and cutting-edge applications of medical imaging AI. The research code is open-sourced on GitHub (https://github.com/enricferrera/tfg-colorectal-cancer-gnn) and was published on June 8, 2026.

Colorectal cancer is the third most common cancer globally, and assessing lymph node metastasis risk in pT1-stage patients is a clinical decision-making challenge. Traditional pathology relies on manual observation, which struggles to capture spatial relationships between cells; CNNs focus on local textures and cannot explicitly model topological relationships. GNNs treat cells as nodes and spatial proximity as edges, preserving geometric structures while learning high-order features, providing a new approach to solving this problem.

The study developed a graph generation module supporting four adjacency relationship strategies: Euclidean k-NN graph (local communities), cosine similarity k-NN graph (functional clustering), radius graph (interaction range), and fully connected graph (global information). Cell-level features (CLS tokens) are extracted via deep learning and integrated with clinical metadata to link micro-level features with macro-level outcomes. Both GPU-accelerated and CPU-parallel versions are provided to balance performance and compatibility.

The project implements multiple targeted GNN architectures:

1. GATWeight_batch: An improved GAT that integrates edge similarity into message passing to distinguish between types of intercellular interactions;
2. GCNWithAgg: A standard GCN with global mean pooling to aggregate node-level representations into patient-level results;
3. Hierarchical TopKPooling: Iteratively selects important nodes to build a coarsened graph, simulating pathologists' hierarchical observation to identify cross-scale patterns.

To address class imbalance in medical data, loss weights are calculated via class_weights.py to focus on minority classes; stratified 10-fold cross-validation is used to ensure consistent sample proportions; MLflow is integrated to track hyperparameters and model versions; mixed-precision training is used to accelerate the process and fully utilize GPU resources.

The project developed the t_SNE.py tool to generate projection maps, showing class clustering in feature space, cell distribution in different tumor regions, and the model's ability to distinguish tissue patterns between metastatic and non-metastatic patients. Interpretability analysis is crucial for medical AI, helping clinicians understand the basis for decisions.

This study represents a cutting-edge direction in computational pathology, shifting towards modeling cellular spatial relationships and microenvironments. If clinically validated, it can assist in pathological diagnosis, optimize treatment plans, and screen clinical trial participants. However, large-scale multi-center validation, regulatory approval, and workflow integration are required, and the technical path provides a reference for the field.