Satoshi Kasai

The purpose of this paper is to propose a detection scheme for masses existing around thick mammary gland regions. The scheme includes a template-matching technique with four reference patterns, which are partial images extracted from a Gaussian distribution, and the feature values determined by concentrating feature and density gradient. The new algorithm consists of 11 steps: (i) image digitization, (ii) extraction of breast region, (iii) reduction of image matrix, (iv) dynamic-range compression, (v) density gradient calculation, (vi) extraction of pectralis muscle region, (vii) overall detection, (viii) elimination of false positives (1), (ix) regional detection, (x) elimination of false positives (2) and (xi) indication of detected masses. Although stage (ix) made it possible to detect the masses existing around thick mammary gland regions, the number of false positives on this region increased. Stage (x) was added for the elimination of new false positives that were detected by stage (ix). A total of 2, 008 digitized mammograms was used for the performance study. As a result, our new scheme identified 95% of the true masses with 2.4 false positives per image. It was possible to detect 43 of 52 masses that were not detected by our previous method. These results indicate that this proposed method is effective, although the process for elimination of false positives should be improved.