The purpose of this study is to improve our automated detection algorithm for masses on digital mammograms by using eleven kinds of features, because the elimination of false positives (FPs) consisting of parts of mammary glands has been required in our mammogram computer-aided diagnosis(CAD) system. These features are (1) length-to-width ratio, (2) minimum width, (3) circularity, (4) average contrast in the candidate, (5) average contrast in the central part of the candidate, (6) average of the standard deviations of pixel-value distributions in the equal distance from the gravity, (7) roughness(+) in the central part of the candidate, (8) roughness(-) in the central part of the candidate, (9) percentage of the gradient-component ratio in constant directions determined by nipple position, (10) standard deviation for the central part of the candidate in unsharp-mask processed image, and (11) gradient ratio for each direction obtained using the gravity. The discriminant analysis applied to these 11 features was employed to eliminate the FPs. The sensitivity of our revised algorithm was 87% with 1.7 FP detections per image in our database of 884 digitized mammograms, which demonstrates the effectiveness of our proposed method for eliminating FPs of funicular-shaped structures.

View full abstract