A GNN Architecture for Integrating Multi-Tissue WSIs in Ovarian Cancer

This thesis explores the challenge of survival prediction in oncology by leveraging histological images from multiple tissue types. The aim is to construct a model that effectively integrates multi-tissue information using Graph Neural Networks (GNNs). Starting from Whole Slide Images (WSIs), vector embeddings are extracted using the TITAN method, and patient-specific graphs are constructed to represent inter-tissue relationships. The model is validated on a real dataset using 5-fold cross-validation with grid search. The results demonstrate an improvement in predictive performance compared to single-tissue models, with promising accuracy and c-index values. Additionally, attention score analysis offers interpretable insights into which tissues are most clinically informative. These findings suggest that structured multi-tissue integration via graph neural networks (GNNs) is a promising approach for computational medicine.