Computational pathology algorithms for nasopharyngeal carcinoma prognosis

Recent advances in artificial intelligence have shown great promise in cancer research; however, their application to rare cancers such as nasopharyngeal carcinoma (NPC) remains relatively underexplored. This thesis presents a series of computational approaches for NPC prognosis using Haematoxylin and Eosin (H&E)-stained whole-slide images (WSIs). NPC is a rare malignancy characterised by strong immune activity and driven in part by Epstein–Barr virus (EBV) infection. Current diagnostic and prognostic tools are often limited by subjectivity, high cost, and poor generalisability, particularly in cases of local recurrence and distant metastasis. By leveraging deep learning, multiple instance learning (MIL), and graph neural networks (GNNs), this work explores how routinely available histology data can support risk strati!cation and enhance biological understanding. We introduce a multi-scale MIL framework for predicting the expression of LMP%, an EBV-encoded oncogenic protein, directly from H&E-stainedWSIs. This approach captures both !ne-grained cellular features and broader tissue context, o-ering a scalable alternative to immunohistochemistry. We then shift our focus on quantifying tumour-in!ltrating lymphocytes (TILs) through spatial analysis of tumour and immune nuclei. These features were used to construct a digital immune-based risk score, which demonstrated improved prognostic value for local recurrence compared to traditional clinical markers. Finally, we present a graph-based deep learning model that integrates morphological and clinical features to predict the risk of distant metastasis. This model achieved strong prognostic performance and revealed associations with tumour aberrations and immune activity when validated across multiple cohorts. Collectively, our !ndings highlight the potential of computational pathology to deliver scalable, interpretable, and biologically meaningful tools for improving clinical decision-making in NPC.