Applications of digital pathology for the analysis of high grade serous ovarian cancer

Abstract

Background and Aims Ovarian cancer is one of the most common gynaecological malignancies in women and the most lethal. High grade serous ovarian cancer (HGSOC) accounts for the majority of ovarian cancer deaths. Expert pathologists play a critical role in the diagnosis of ovarian cancer and in the stratification of patients for certain therapeutics. However, visual analysis can be subjective and this can result in patients potentially missing out on treatment options. Pathology services have come under increasing pressure in recent years. Cancer incidence is rising and, with an aging population, will continue to do so. Although the introduction of new diagnostic tests has brought some much needed improvements to cancer survival rates, they have further increased the demand on pathology services. The Covid-19 pandemic has also added to the strain, with a growing backlog of diagnoses over time. This thesis explores some applications of digital pathology in the analysis of HGSOC and aims to determine if automated methods can offer a reliable solution to the aforementioned clinical and operational challenges. Methods Formalin fixed paraffin embedded (FFPE) HGSOC samples from the Northern Ireland Biobank, Edinburgh Ovarian Cancer Database and Aberdeen University were used in the investigations in this thesis. Digital pathology methods using commercial software, TissueMark™ and open source software, QuPath, were used to identify, classify and quantify tumour and immune cells. These results were compared to visual assessments made by expert pathologists and manual counts to assess accuracy. The prognostic significance of known biomarkers was investigated and new methods of assessment and scoring were proposed for HGSOC. Results Investigation of the performance of a commercial product, TissueMark™ revealed a lower level of accuracy compared to an expert pathologist. However, TissueMark™ could be used as a tool to “triage” hematoxylin and eosin (H&E) samples prior to macrodissection. A cohort of 60 SOC patients was stained for PD-L1 using two leading commercial antibodies and then analysed and scored both by an expert pathologist and using a digital pathology approach. In total, 240 individual analyses were performed. A method that successfully identified the prognostic value of PD-L1 was established and this was then further investigated and validated using an automated approach in a HGSOC-specific cohort of 81 patients. Our results indicate that digital pathology can be a useful addition to the scoring of these biomarkers. Further investigation of the tumour microenvironment of SOC and HGSOC using QuPath was carried out on a cohort of 71 SOC patients and on a total of 299 HGSOC patients from two separate cohorts for three different markers. This total of 1110 individual analyses confirmed the prognostic significance of cluster of differentiation 3 (CD3) and cluster of differentiation 8 (CD8) and demonstrated the reliability of digital pathology in analysis of immunohistochemistry (IHC) for immune markers. Finally, we investigated the transcriptomic profile of 218 HGSOC patients alongside an automated IHC analysis approach using QuPath, to identify T-cell immune response. We found significant transcriptional differences between certain genes in the CD3 and CD8 High and Low groups. These included C-C motif chemokine ligand 5 (CCL5), C-C motif chemokine receptor 5 (CCR5) and lymphocyte-specific protein tyrosine kinase (LCK), which have been established as therapeutic targets in other cancer types. We also identified Perforin 1 (PRF1) as another potential indicator of prognosis. Conclusion This thesis has demonstrated that digital pathology can offer fast and reliable quantification of tumour and immune cells in HGSOC and can successfully identify prognostic significance of key markers using specified protocols. Automated scoring and analysis could help to relieve some of the mounting pressure on pathology services. However, the interpretation of microscopic images is a holistic process, requiring extensive clinical knowledge and experience and it is unlikely that a fully automated approach to pathology is achievable.