Dependence of computer classification of clustered microcalcifications on the correct detection of microcalcifications

Our purpose was to study the dependence of computer performance in classify ing clustered microcalcifications as malignant or benign on the correct det ection of microcalcifications. Specifically, we studied the effects of comp uter-detected true-positive microcalcifications and computer-detected false -positive microcalcifications in true microcalcification clusters. Using a database of 100 mammograms, we compared computer classification performance obtained from computer-detected microcalcifications to (1) computer classi fication performance obtained from manually identified microcalcifications, and (2) radiologists' performance. When an artificial neural network (ANN) was trained with manually identified microcalcifications, computer classif ication performance was comparable to or better than radiologists' performa nce as the number of computer-detected true-positive microcalcifications de creased to 40% and as the number of computer-detected false-positive microc alcifications increased to 50%. Further loss in computer-detected true-posi tive microcalcifications degraded classification performance substantially. Moreover, training the ANN with computer-detected microcalcifications also degraded computer classification performance. These results show that comp uter performance in classifying clustered microcalcifications as malignant or benign is insensitive to moderate decreases in computer-detected true-po sitive microcalcifications and moderate increases in computer-detected fals e-positive microcalcifications. (C) 2001 American Association of Physicists in Medicine.