Pathology-prior driven substructure-aware graph neural network for whole slide image classification

Abstract

Accurate detection of sentinel lymph node metastasis in whole slide images (WSIs) is critical for breast cancer staging, yet existing graph neural network (GNN) based methods lack pathological prior features that characterize diagnostically meaningful tissue configurations, leaving the model without an effective guidance for message passing or node importance evaluation. We propose a pathology-prior driven substructure-aware graph neural network for whole slide image classification. The pathology-prior substructure encoding (PPSE) module characterizes four substructure descriptors. The substructure-aware message passing (SAMP) module uses these descriptors to guide message passing via edge-level attention weights computed from incident node substructure features, concentrating information flow on diagnostically coherent connections. The substructure-aware graph readout (SAGR) module uses substructure-enriched node representations to evaluate node importance by deviation from the slide-level tissue distribution, shifting node scoring from absolute semantic evaluation to slide-level anomalousness detection. Experiments on the SLN-Breast dataset demonstrate AUC of 0.9479±0.0133 and ACC of 0.9462±0.0188, outperforming most baselines with substantially reduced cross-fold variance. Independent validation on the TCGA-ESCA dataset further demonstrates AUC of 0.9610±0.0295 and ACC of 0.9050±0.0284, confirming the generalizability of the proposed framework across cancer types. Code is available at https://github.com/jyatgithub/Pathology-Prior-Driven-Substructure-Aware-Graph-Neural-Network-for-WSI.