Artificial Intelligence for Subgrouping of Non-Hodgkin Lymphoma

Non-Hodgkin Lymphoma (nHL) is a common type of cancer, accounting for over 10,000 annual diagnoses in England. Diagnostic procedures for nHL require analysis of tissue biopsies by highly skilled histopathologists. Procedural demand has increasingly outpaced supply, with potential for negative impact on patient care. In this thesis, methods of automating high burden, subjective procedures and of extracting additional value from routine samples are proposed, with a focus on the most common forms of nHL: Follicular Lymphoma (FL) and DLBCL.

A pipeline of processes are proposed for automation of FL cytological grading, involving quantification of cancerous B-cells (centroblasts). Two novel methods of clustering FL and DLBCL patients on cell populations are presented. Finally, FL and DLBCL are grouped on presence of informative proteins (immunohistochemistry).

High cytological grade was found to associate with increased survival risk but, the automated procedure classified a group of FL patients with good outcome. Cell-based clusters associated with mutation in FL and survival in DLBCL. Automated immunohistochemistry estimates clustered patients on gene expression profiles, informative of survival. Prediction of immunohistochemistry clusters using routine (H&E) biopsy slides was successful with state-of-the-art pathology image representations.

Classification of a low risk patient group through automated cytological grading evidenced potential utility of centroblast subtypes as indicators of good outcome in FL. Performance of automated immunohistochemistry quantification in clustering gene expression groups evidenced utility of automated procedures for reduction of histopathologist burden. Successful classification of immunohistochemistry with H&E evidenced feasibility, by extension, of gene expression prediction.