The performance of a CAD system for subtle lesions is generally much lower than their performance for less subtle lesions. The goal of this project is to develop a CAD system using advanced computer vision techniques aiming at improved detection of retrospectively seen cancers on prior mammograms and incorporate the developed CAD system into our current CAD system. During the project years, we have performed the following tasks: (1) collect the data sets of digitized film mammograms for training and testing our CAD system, (2) develop a series of single-view computer vision techniques for mass detection and classification in prior mammograms, (3) reduce FPs by correlation of image information from multiple view mammograms of the same patient, (4) develop a information fusion scheme to combine the new CAD system with the existing CAD system for mass detection, and (5) evaluate the effects of the newly developed CAD scheme with a large data set. We have found that our new computer-vision techniques can significantly improve the performance of the CAD system for mass detection by JAFROC analysis. The significance of this project is that the newly developed CAD system may be able to aid radiologists in detecting breast cancers at an early stage. Since early detection and treatment can reduce breast cancer mortality rate and health care costs, the proposed CAD system will improve the efficacy of mammography for breast cancer screening.
