Hierarchical Graph Neural Networks for Digital Pathology

Gastric cancer is the fifth most common cancer and fourth most common cause of cancer-related death globally. The highest incidence is seen in Eastern Asian countries including Japan, Mongolia and South Korea. Patients in these countries diagnosed with locally advanced resectable (TNM stage II/III) disease are treated with surgery followed by adjuvant cytotoxic chemotherapy according to current Eastern Asia treatment guidelines. However, only a subset of patients benefits from adjuvant chemotherapy and there are currently no methods used in clinical practice to identify these patients.

The work in this thesis presents different graph-based analysis techniques applied to routine Haematoxylin/Eosin stained histopathology resections from 547 gastric cancer patients recruited to the Korean CLASSIC trial. The aim of these analyses were to identify patients who might benefit from adjuvant chemotherapy. This thesis presents a novel method for constructing hierarchy within graphs that enables graph neural networks to use cell and tissue-level information in segmentation and risk-prediction tasks. Using these methods, a potential new biomarker was identified and a risk-prediction algorithm was developed that was able to predict which patients benefit from adjuvant chemotherapy. In addition, attention weights within the risk-prediction algorithm allowed for the discovery of spatially relevant features associated with risk which validated the potential new biomarker and could inform future therapies.