Breast cancer histopathology image analysis

Pathology labs are currently undergoing a transformation towards a fully digital workflow. In addition to the digital management of tissue samples, pathology orders and reports, this includes the digitization of histopathology slides and use of computer monitors for viewing them, which aims to replace the optical microscope as the primary tool used by pathologists. This transformation has only recently been enabled by the introduction of cost and time efficient whole-slide imaging (WSI) scanners. This process of adoption of WSI is somewhat analogous to the digitization of radiological imaging. One of the main benefits of digital slides compared with conventional glass slides is that they enable seamless integration of quantitative automatic image analysis methods into the pathology lab workflow. A relatively large percentage of the samples that are analyzed in pathology labs are from breast cancer patients, since this disease is the most prevalent form of cancer among women. Analysis methods that are routinely performed by pathologists, such as determination of the histological grade and the hormone receptor status by immunohistochemistry, can be tedious and are hampered by observer variability. The histological tumor grade is commonly determined according to the modified Bloom-Richardson system, which consists of semi-quantitative assessment of nuclear atypia, tubule formation and mitotic activity in hematoxylin and eosin (H&E) stained sections. The analysis of immunohistochemically stained slides mainly involves the estimation of the number of cells that are positive for a particular antigen and the degree of positivity (staining intensity). The focus of this thesis is on automatic image analysis of H&E stained breast cancer histopathology images. First, a detailed review of the literature on the topic of breast cancer histopathology image analysis is presented. The tissue preparation and imaging processes are also covered and particular attention is given to techniques for detection and segmentation of various objects, such as nuclei, tubules and mitotic figures, as well as computer-aided diagnosis and prognosis methods. Then, the development and evaluation of a method for automatic segmentation of nuclei in H&E stained breast cancer histopathology images is presented. The proposed nuclei segmentation method is further used to extract and evaluate the prognostic value of nuclear morphometric features in a cohort of 101 male breast cancer patients. The remainder of the thesis deals with the difficult problem of mitotic figures detection in H&E stained slides. It describes the development and evaluation of a mitotic figures detection method and presents the results from a challenge workshop on this topic.