Master Thesis - Automatic Malignancy Detection in Breast Histopathology Images - Siemens Corporate Research, Princeton, NJ and TU Munich, Germany
Abstract: In traditional cancer diagnoses, pathologists examine biopsies to make diagnostic assessments largely based on deviations in the cell structures and changes in the cell distribution across the tissue. Detection of malignancy from histopathological images of breast cancer is a labor-intensive and error-prone process. Automation of this process is highly desirable and would streamline the clin- ical workflow. In this work, we present an efficient Computer Aided Diagnosis system that can differentiate between cancerous and noncancerous H&E (hemotoxylin&eosin) biopsy samples. The proposed approach aims at extracting relevant features that are used to quantify the observed changes in cancer tissue. We explore texton-based, network-based and novel morphometric features that take advantage of the special shape of the nuclei cells in breast cancer histopathological images. In extracting the network-based and morphometric features, the exact locations and segmentation of the nuclei needs to be deter- mined, which is accomplished using a generalized fast radial symmetry transform and the Random Walker algorithm, respectively. The proposed approach also uses Urquhart graph generation that can better discriminate between different anatomical structures being an approximation of the relative neighborhood graphs, which are known for matching human perceptions of the shape of a given set of points. Using a Support Vector Machine classifier on these features in conjunction with specific feature selection techniques, we differentiate between malignant and benign breast cancer histological slides. Experiments were conducted using H& E stained samples, previously annotated by pathologists. Our method achieved a high average sensitivity and specificity.