Studies on detection and analysis of mass shadows in mammography (in Japanese)

Abstract

Breast cancer is increasing rapidly in Japan, intensifying the need for a computer-aided diagnosis (CAD) system to help physicians interpret mammograms more efficiently. Reported here is an attempt to improve the accuracy of mass detection algorithms for mammograms in a CAD system currently under development. This CAD system uses multiresolution processing for pre-processing in which images are decomposed into plural frequency bands, with each frequency band enhanced independently. The algorithm enhances masses while simultaneously suppressing such low-frequency components as background trend and such high-frequency components as mammary glands. Two newly developed methods of detecting abnormalities are described, one based on an adaptive thresholding technique and one based on similarity to a funnel-shaped model. The former applies to mass shadows isolated from normal regions, while the latter applies to mass shadows that appear on the peripheries of thick mammary gland regions. In addition, three methods of eliminating false-positive candidates have been developed: discrimination based on second-order statistics calculated from co-occurrence matrix and gray-level difference methods, comparison of left and right breast images, and recognition of funicular-shaped false-positive candidates through the use of discriminant analysis. In summary, the new algorithms described here have significantly improved the overall performance of the CAD system involved.