AbstractBreast cancer is currently the most common cancer among women worldwide.
Mammography has been the most reliable and effective screening tool for the early detection of breast cancer. Recently, computer-aided diagnosis has become a major research topic in medical imaging and has been widely applied in clinical situations. This thesis investigates the employment of computer vision and image processing techniques for mammographic image analysis, which focuses on the aspects of mammographic risk assessment and microcalcification classification. Many studies have indicated that mammographic density and mammographic parenchymal patterns are both strong predictive markers of breast cancer risk in mammographic images, which play an important role in estimating breast cancer risk. In this thesis, we present a variety of methods for estimating mammographic density and modelling mammographic parenchymal patterns. In order to build a complete framework for automated mammographic risk assessment, we first develop a breast region segmentation method as a pre-processing step, which segments the breast region from mammograms to provide fundamental data for subsequent analysis. Subsequently, we propose two breast density segmentation methods to estimate mammographic density. The rst method is based on a modified fuzzy c-means algorithm which incorporates spatial information into the classic fuzzy c-means algorithm. The breast region is segmented into a number of sub-regions corresponding to di erent densities by clustering pixels with similar greylevel values. The second method exploits a topographic map to represent the overall pro le of breast tissue density within the breast. Dense tissue regions are segmented by detecting prominent/independent shapes based on a shape tree. For modelling mammographic parenchymal patterns, we present a method to model breast tissue appearance based on statistical analysis of local tissue appearance. Five different types of local features are investigated, covering the aspects of intensity, texture, and geometric structure. In addition, a multiscale blob based representation is proposed to model mammographic parenchymal patterns. Instead of statistically describing breast tissue appearance within the whole breast, we focus only on dense tissue with approximately blob-like structures. The validity of the proposed methods is evaluated using the MIAS and DDSM databases. A high agreement with expert radiologists is indicated according to the BIRADS density classi cation. The obtained classi cation accuracies are up to 79.44% and 81.23%, and increase to 93.15% and 91.70% for the low/high risk classi cation. For microcalci cation classi cation, a novel method is developed based on topological analysis. The connectivity between individual microcalci cations is analysed to classify microcalci cation clusters into malignant and benign. This method is evaluated using three datasets: MIAS, DDSM, and a full- eld digital dataset. High classi cation accuracies (up to 96%) and good ROC results (area under the ROC curve up to 0.97) are achieved.
|Date of Award||03 Jun 2013|
|Supervisor||Reyer Zwiggelaar (Supervisor) & Frédéric Labrosse (Supervisor)|