Computer aided detection of microcalcification clusters in digital mammogram images. (2004)
Type of ContentTheses / Dissertations
Thesis DisciplineElectrical Engineering
Degree NameDoctor of Philosophy
PublisherUniversity of Canterbury. Electrical and Computer Engineering
AuthorsMcLoughlin, Kirstin J.show all
Recent advancements in computer technology have ensured that early detection of breast cancer, via computer aided detection (CAD) schemes, has become a rapidly expanding field of research. There is a desire to improve the detection accuracy of breast cancer without increasing the number of falsely identified cancers. The CAD scheme considered here is intended to assist radiologists in the detection of micro calcification clusters, providing a real contribution to the mammography screening process. Factors that affect the detection accuracy of micro calcifications in digital mammograms include the presence of high spatial frequency noise, and locally linear high intensity structures known as curvilinear structures (CLS). The two issues considered are how to compensate for the high frequency image noise and how to detect CLS thus removing their influence on micro calcification detection. First, an adaptive approach to modelling the image noise is adopted. This is derived directly from each mammogram and is adaptable to varying imaging conditions. It is found that compensating for the high frequency image noise significantly improves micro calcification detection accuracy. Second, due to the varying size and orientation of CLS in mammogram images, a shape parameter is designed for their detection using a multiresolution wavelet filter bank. The shape parameter leads to an efficient way of distinguishing curvilinear structures from faint micro calcifications. This improves micro calcification detection performance by reducing the number of false positive detections related to CLS. The detection and segmentation of micro calcification clusters is achieved by the development of a stochastic model, which classifies individual pixels within a mammogram into separate classes based on Bayesian decision theory. Both the high frequency noise model and CLS shape parameters are used as input to this segmentation process. The CAD scheme is specifically designed to be independent of the modality used, simultaneously exploiting the image data and prior knowledge available for micro calcification detection. A new hybrid clustering scheme enables the distinction between individual and clustered micro calcifications, where clustered micro calcifications are considered more clinically suspicious. The scheme utilises the observed properties of genuine clusters (such as a uniform distribution) providing a practical approach to the clustering process. The results obtained are encouraging with a high percentage of genuine clusters detected at the expense of very few false positive detections. An extensive performance evaluation of the CAD scheme helps determine the accuracy of the system and hence the potential contribution to the mammography screening process. Comparing the CAD scheme developed with previously developed micro calcification detection schemes shows that the performance of this method is highly competitive. The best results presented here give a sensitivity of 91% at an average false positive detection rate of 0.8 false positives per image.