Computer Aided Diagnosis of Breast Cancer for Diverse Mammographic Densities

Abstract

Breast cancer is a major public health problem affecting millions of women worldwide. Mammography is the most widely employed screening modality for the early detection of breast cancer. However, the sensitivity of a mammogram is affected by the breast tissue density. This is due to variations in parenchymal patterns and intensities across various tissue types with X-ray absorption. Computer aided detection/diagnosis (CAD) systems can aid radiologists in better newlinediagnosis of breast cancer by performing an objective assessment. Most of the researchers have proposed CAD systems that have overlooked dense cases and concentrated only on low breast density categories wherein the analysis is relatively straightforward as the abnormalities are clearly distinguishable from the fatty background. This thesis focuses on developing a CAD system, targeting mass detection and classification for breast cancer diagnosis, considering all density categories including dense breasts. The performance of false positive reduction associated with mass detection and that of mass classification are affected by the density of the breast tissue, due to factors such as obscuring of mass boundaries by surrounding normal tissues, and similarity in appearance of normal dense and masses. These factors result in segmentation inaccuracies, which have a cascaded effect on feature extraction and classification performance. This thesis explores methods devoid of segmentation and global representation, by employing local characterization of masses to improve the accuracy of mass detection, specifically in dense breasts. Further, an ensemble framework based on segmentation independent features comprising individual learning frameworks each specialized for a particular density category is also explored to alleviate the issues in modeling all density categories using a single learning framework. In addition, to further enhance the performance of mass characterization, a hybrid approach that combines global and local analysis is employed to analyze low..