Computer-aided diagnoses of breast cancer

Breast cancer is the most common type of cancer among women in the western world. Mammography is the most important imaging modality for the early detection and the diagnosis of breast cancer. However, the characterization of breast lesions as benign or malignant, based on their appearance in mammograms, is a difficult task. It is reported that usually less than 30% of all breast biopsies actually show a malignant pathology. The high number of unnecessary breast biopsies causes major mental and physical discomfort for the patients as well as unnecessary expenses. In recent years, there was considerable interest in the development of computer-aided diagnosis (CADx) systems that aim to support radiologists in the discrimination of benign and malignant mammographic lesions. However, so far, no mammography CADx system is available commercially. The goal of this thesis is the design and implementation of novel systems for the computer-aided discrimination of benign and malignant mammographic lesions. While a transparent decision process is often considered as an important requirement for the acceptance of CADx systems in clinical practice, the decision process in state of the art CADx approaches is usually not intelligible to radiologists but acts like a black box. Hence, the focus of this work is on the design and implementation of CADx systems with decision processes that are transparent to radiologists. This goal is met by using knowledge-based reasoning systems to classify mammographic lesions. Additional goals of this work are the development and implementation of novel approaches to the sub-problems of CADx systems: lesion segmentation, feature extraction, feature selection, feature weighting and classification. Novel algorithms for these problems are presented, and their performance is compared to state of the art approaches on clearly defined and publicly available data sets. In the context of lesion segmentation, a novel approach for the segmentation of clustered microcalcification and two novel approaches for the segmentation of masses are proposed. New feature extraction approaches are proposed, and implemented for masses and calcifications. In addition, a large set of state of the art features is implemented and evaluated in a systematic manner. Three methods for the automatic selection of optimal feature subsets are proposed and compared. Finally, knowledge-based and decision-tree approaches to the classification of lesions are investigated and both single- as well as multi-view CADx systems are built and evaluated.