Spatial context-aware breast cancer molecular subtype prediction in whole-slide images using graph convolutional networks: Cross-domain evaluation

Breast cancer molecular subtype determination is crucial for personalized treatment planning but currently relies on expensive gene expression profiling or immunohistochemistry with discordance rates up to 30%. This study presents a systematic comparison of context-awareness mechanisms for molecular subtype prediction directly from H&E whole-slide images, evaluating graph-based spatial modeling against attention-based aggregation across multiple architectures, datasets, and deployment scenarios.

The WSI2Graph framework is adapted to represent whole-slide images as spatial KNN graphs, enabling a controlled comparison of context-aware (CA) graph convolutional aggregation against non-context-aware (NCA) attention-based pooling under identical feature extraction. Three evaluation protocols (EP1-EP3) assess within-domain performance (EP1; BCNB dataset, 1058 core-needle biopsies), constrained cross-domain transfer (EP2; SBC dataset, 533 surgical resection specimens), and foundation model integration (EP3; CONCH), all validated with McNemar's paired tests. Graph topology characterization, quantitative interpretability analysis along three axes (attention concentration, tissue composition, spatial coverage), and per-class clinical metrics for therapeutically relevant subtypes are additionally reported.

Within-domain (EP1), CA significantly outperforms NCA on all three classification tasks (p<0.025), with accuracy gains of +4.6, +9.2, and +8.7 percentage points for binary, ternary, and quaternary subtype classification. A capacity and coordinate ablation shows that aggregator capacity accounts for a substantial portion of the observed accuracy advantage, while true spatial coordinates reduce training variance across tasks. Under constrained transfer (EP2), CA matches NCA on binary and quaternary tasks, with only the ternary task showing a significant NCA advantage (p=0.002). With foundation model features (EP3), no significant CA-NCA differences were detected (p>0.07), consistent with feature quality and the broader retraining scope of EP3 outweighing aggregation strategy for cross-domain transfer. Interpretability analysis shows both architectures enrich for tumor and inflammation, with the top-attended regions of CA spanning 66% of the biopsy area versus 29% for NCA; both progressively converge on the most challenging task.

Context-aware models significantly outperform attention-based aggregation within-domain, but this advantage does not persist uniformly under domain shift. Cross-domain transfer depends on the task, on feature quality together with retraining scope, and is associated with the structural domain shift between core-needle biopsies and surgical resections, rather than being uniform. These findings provide practical guidance for architecture-feature selection across deployment scenarios.