Using Graph Neural Networks to Capture Tumor Spatial Relationships for Lung Adenocarcinoma Recurrence Prediction.

Quantitative histomorphologic features derived from H&E-stained whole slide images (WSIs) have improved patient outcome prediction in non-small cell lung cancers (NSCLC). However, most existing works utilize tile-based prediction but ignore the spatial relationships between tiles. Given that NSCLC exhibits tissue heterogeneity, capturing these spatial relationships is important in determining tumor aggressiveness. In this work, we predict recurrence-free survival in lung adenocarcinoma (LUAD), the most common subtype of NSCLC, using a graph neural network (GNN) that models a WSI as a graph where each tile is a node. We incorporated important prognostic biomarkers for LUAD, including histologic subtypes and tumor immune microenvironment, into the pipeline. We hypothesize that the GNN model built based on domain-specific features captures the spatial connectivity between tiles in WSIs and results in more accurate recurrence-free survival prediction in LUAD. The effectiveness of our method was demonstrated via comparison with multiple baseline models using 201 publicly available Cancer Genome Atlas (TCGA) LUAD patients.