Hierarchical graph networks for breast cancer subtype classification
TL;DR
A hierarchical Graph Neural Network (GNN) framework for ROI-level breast cancer subtype classification that represents nuclei and tissue regions as linked graph structures is presented and shows that sequential hierarchical fusion is the most effective configuration in this setting.
Abstract
Breast cancer histopathological diagnosis relies on the expert assessment of Hematoxylin and Eosin (H&E) stained slides. This process is both time-consuming and prone to inter-observer variability. Deep learning methods have shown great potential for automating cellular analysis; however, conventional patch-based approaches do not explicitly model relational spatial dependencies between biological entities, potentially limiting structural interpretability. We present a hierarchical Graph Neural Network (GNN) framework for ROI-level breast cancer subtype classification that represents nuclei and tissue regions as linked graph structures. The study provides a validation-driven design-space analysis of graph topology, node features, spatial pooling, and fusion of cell and tissue information within this hierarchical pipeline. The proposed approach was evaluated on the BRACS dataset, which comprises 547 whole-slide images and over 4500 annotated regions spanning seven histopathological classes. Results show that sequential hierarchical fusion is the most effective configuration in this setting, outperforming alternative strategies for integrating cell and tissue information and yielding particularly consistent improvements for Atypical Ductal Hyperplasia, Flat Epithelial Atypia, and Ductal Carcinoma in situ.
