The goal of the proposed research is to develop computer-based techniques in order to improve diagnostic accuracy of breast cancer detection in mammography. Specifically, we plan to (1) develop an automated computer scheme for the detection and characterization of microcalcifications in digital mammograms, and (2) develop new digital image processing techniques to improve the visual detectability of abnormalities in mammograms. The automated detection scheme proposed will aid radiologists in screening mammograms for microcalcifications, thereby reducing the miss rate due to human error. The proposed image processing methods will be tailored to suit the characteristics of mammographic images and the image display system, thereby improving the contrast sensitivity for subtle mammographic features in the displayed image. The physical characteristics of microcalcifications will be analyzed in a data base of clinical mammograms. Based on these characteristics, we will develop efficient spatial filtering methods that can effectively remove the structured background from the mammogram. Signal-extraction techniques will then be designed to isolate microcalcifications from the remaining noise background. To facilitate the evaluation of the effectiveness of various image-processing and signal-extraction techniques, Monte Carlo methods will be employed to generate simulated microcalcifications which can be superimposed at known locations on mammograms. An observer performance study will be conducted to evaluate the effects of the computer-aided method on radiologists' performance. A pilot study will also be performed to characterize the features of microcalcifications that are associated with malignant and benign processes. We will develop new image processing methods which combine a dynamic range compression technique with overall contrast enhancement. This new approach will optimally utilize the dynamic range of an image display system and increase the perceived signal-to-noise ratio of mammographic lesions. We will first evaluate theoretically the effects of various image processing parameters on visual detectability by using a perceived statistical decision theory model of signal-to-noise ratio. Image processing methods selected on the basis of these theoretical studies will then be applied to clinical mammograms, and the improvement in diagnostic accuracy will be determined by observer performance studies.
Showing the most recent 10 out of 36 publications
