Project 466562

Ovarian cancer ranks fifth in cancer deaths among women (~1,800 and ~14,000deaths/year in Canada and U.S., respectively), accounting for more deaths than any other cancer of the female reproductive system. There are 5 major subtypes of epithelial ovarian cancer. Each subtype is associated with a distinct set of mutated genes, cellular structures, and responds differently to chemotherapy. However, their treatment remains mostly the same, centered around surgery and chemotherapy. The lack of subtype specific treatment can be attributed to the challenges in sub-classifying ovarian cancer. One major challenge in subclassifying ovarian cancer is poor diagnosis reproducibility.Histopathology (the microscopic study of diseased tissue) has been an important tool incancer diagnosis. Histopathology images capture cellular phenotypes which reflect the effect of DNA alterations on cancer cells. However, sub-classification based on histopatological imagesis often hampered by disagreement among independent observers and poor diagnostic reproducibility. To address this issue,I intend to build a novel artificial intelligence (AI)-based platform for an automated analysis of histopathology slides and ovarian cancer biomarker discovery.Recent applications of deep learning methods (e.g. convoluted neural networks (CNNs)) on image data have demonstrated the ability of CNN to learn predictive features from raw imaging data.I will explore different architectures of CNNs to accurately identify the subtypes of ovarian cancer and choose the best-performing classifer by measuring its performance on held-out test datasets.I expect this research to contribute to the identification of ovarian cancer biomarkers for patient stratification and individualized treatment.