Development of Deep Learning Based Automated Methods for Breast Cancer Histopathology Image Analysis

Abstract

According to the recent report by the Global Cancer Observatory, breast cancer has overtaken lung cancer as the leading type of cancer in terms of new cases reported. In 2020, breast cancer accounted for 11.7% of all new cancer cases and 6.9% all cancer related deaths. Timely diagnosis and targeted treatment can significantly improve the survival chances of breast cancer patients. Pathological procedures are integral parts of cancer diagnosis and treatment planning. In the routine cancer pathology analysis, tissue samples are extracted from the tumor regions and applied with suitable stain- ing agents. The glass slides prepared this way are analyzed by pathologists through a microscope to make interpretations about the disease condition. The manual proce- dure of microscopy analysis is tedious, time consuming, and error-prone. Digitization of pathological glass slides into slide images opens a plethora of possibilities to apply computational methods to automate several pathology procedures. The focus of this the- sis work is to develop computational methods for automated analysis of breast cancer histopathology images and extract clinically relevant information to support prognosis and treatment planning. Grading and molecular subtyping of breast cancer are the two important pathology procedures considered for automation in this thesis work. Partic- ularly, automation of two breast cancer grading procedures namely mitosis detection and nuclear atypia scoring are taken as the first two objectives. The third objective is automated molecular subtyping of breast cancer, a classification that supports targeted treatment and hence better outcome. Breast cancer grading categorizes the disease based on its aggressiveness. The grade information is used for prognosis and treatment planning. Among the three pa- rameters involved in breast cancer grading (mitotic cell count, nuclear atypia score, and 6 tubule formation), mitotic cell counting is the most challenging task for pathologists. It is possible to automate this task by applying computational algorithms on pathol- ogy slide images. Lack of sufficiently large datasets, and class imbalance between mitotic and non-mitotic cells are the two major challenges in developing effective deep learning-based methods for automated mitosis detection. In order to address these chal- lenges, an approach of combining datasets from different sources and a more effective image data augmentation technique are used. Following these, a novel method pipeline is proposed which makes use of an advanced deep learning algorithm to address this problem. In contrast to the existing methods that are trained and validated on inde- pendent datasets, the proposed approach aims to develop generalized dataset-agnostic solutions for mitosis detection. The results obtained for the proposed method show improvement over existing deep learning methods based on independent datasets. Nuclear atypia score is the second parameter used for grading breast cancer. Manual procedure of nuclear atypia scoring is laborious and marked by pathologists’ disagree- ment as well as low reproducibility. Automation of this procedure using computational methods is seen as a viable alternative to these challenges. It is observed that most of the existing methods rely on extracted feature-based learning algorithms. Deep learn- ing algorithms are not sufficiently utilized to address this task. In this thesis, a novel deep learning based framework for automated nuclear atypia scoring of breast can- cer is proposed. The framework consists of three major phases namely preprocessing, deep learning, and postprocessing. In the proposed approach, the original three-class problem of slide level atypia scoring is reformulated as a six-class problem of nuclei classification for the effective use of deep learning algorithms. The method based on this framework gives a performance that exceeds the state-of-the-art by a significant margin. Molecular subtyping classifies cancer based on the expression of genetic alterations behind the disease. Identifying the specific subtype aids in targeted treatment of the disease to achieve better outcome. Molecular subtyping through immunohistochem- 7 istry (IHC) analysis is a pathology procedure to determine the subtype of breast can- cer. The existing manual procedure involves assessing the status of the four molecular biomarkers ER, PR, HER2, and Ki67. To automate this procedure, a deep learning- based framework using IHC image analysis is proposed. At present, there are no meth- ods found in literature for IHC based automated molecular subtyping. The proposed system is evaluated for the performance of individual biomarker status predictions and patient-level subtype classification. The results obtained at the various levels of evalua- tions are highly promising. In the extensive literature study carried in the preliminary stage of the research work, it is understood that the potential of deep learning algorithms is not fully utilized in the automation of pathology procedures for mitosis detection and nuclear atypia scoring. The bottlenecks for this are identified and potential solutions are investigated in this thesis work. The performance of proposed methods for these tasks validates the rele- vance of the solution approach adopted. In the absence of any prior work in the literature for automated molecular subtyping of breast cancer, the proposed deep learning-based classification framework establishes a new direction for automating this labor-intensive pathology procedure. The high performance of the proposed method is a strong indica- tion of the clinical applicability of automated methods. In essence, by automating three key pathology procedures in breast cancer diagnosis and treatment planning, this thesis work aims to contribute to the global research efforts towards making cancer treatment more effective, affordable, and accessible.