Significant feature extraction from whole-slide images for diagnosis and prognosis of triple negative breast cancer

Abstract

Breast cancer, the most commonly diagnosed cancer in women worldwide, presents a significant challenge with approximately one in every three newly diagnosed cancers being breast cancer. The United States alone witnesses around 250,000 new cases annually, and the global count of diagnosed women reached 2.3 million in 2020. Recently, artificial intelligence (AI) and deep learning have emerged as promising tools in the field of Digital and Computational Pathology. They offer transformative capabilities, assisting pathologists in their clinical routine by enhancing diagnostic and prognostic abilities.

Traditionally, Pathology involves the analysis of tumor tissue samples obtained through biopsies, which are then scanned to create digital slides or Whole-Slide Images (WSIs). Once these slides are digitized, AI algorithms can perform tasks such as cell counting, pattern detection, and prediction of risk factors, survival rates, and treatment options.

This thesis focuses on two key aspects: diagnosis and prognosis in breast cancer. At the cellular level, we explore the counting of mitosis. This corresponds to a process where pathologists must manually count dividing nuclei from the hematoxilyn and eosin (H&E) WSIs. This is because proliferation is a very strong biomarker in breast cancer, linked to metastasis and survival. Therefore, the first part of the thesis focuses on automatic mitoses counting and an objective tool for assessing proliferation in WSIs using convolutional neural networks (CNNs) under a weakly-supervised paradigm.

Secondly, this thesis delves into the significance of molecular subtypes of breast cancer. These subtypes display varying levels of aggressiveness, prognosis, and treatment responses. Pathologists are unable to derive the molecular subtype from an H&E-stained WSI, and they recur to expensive gene-expression profiling or immunohistochemistry to determine them. For this reason, we employ context-aware approaches and leverage graph-convolutional networks (GCNs) to classify these molecular subtypes only using H&E-stained WSIs, facilitating personalized treatment strategies for pathologists.

Finally, attention is directed towards prognosis, particularly the prediction of survival and distant metastases. Leveraging the power of deep learning, we propose combining the previously mentioned, automatic mitotic score and the image features extracted from the molecular subtypes to develop models capable of accurately forecasting patient outcomes, including the likelihood of metastatic spread. Such predictions hold immense potential for guiding clinical decisions, enabling early interventions, and improving patient care.

In summary, this thesis explores the integration of deep learning and AI in Digital and Computational Pathology, addressing both the diagnostic aspects of automatic proliferation scoring and molecular subtype prediction, as well as the prognostic aspects of survival and distant metastases prediction. Through this work, we aim to equip healthcare professionals with advanced tools and knowledge to combat breast cancer more effectively.