Abstract
Concerted efforts in genomic studies examining RNA transcription and DNA methylation patterns have revealed profound insights in prognostic ovarian cancer subtypes. On the other hand, abundant histology slides have been generated to date, yet their uses remain very limited and largely qualitative. Our goal is to develop automated histology analysis as an alternative subtyping technology for ovarian cancer that is cost-efficient and does not rely on DNA quality. We developed an automated system for scoring primary tumour sections of 91 late-stage ovarian cancer to identify single cells. We demonstrated high accuracy of our system based on expert pathologists' scores (cancer = 97.1%, stromal = 89.1%) as well as compared to immunohistochemistry scoring (correlation = 0.87). The percentage of stromal cells in all cells is significantly associated with poor overall survival after controlling for clinical parameters including debulking status and age (multivariate analysis p = 0.0021, HR = 2.54, CI = 1.40-4.60) and progression-free survival (multivariate analysis p = 0.022, HR = 1.75, CI = 1.09-2.82). We demonstrate how automated image analysis enables objective quantification of microenvironmental composition of ovarian tumours. Our analysis reveals a strong effect of the tumour microenvironment on ovarian cancer progression and highlights the potential of therapeutic interventions that target the stromal compartment or cancer-stroma signalling in the stroma-high, late-stage ovarian cancer subset.