CLASSIFICATION OF MASS AND NONMASS REGIONS ON MAMMOGRAMS USING ARTIFICIAL NEURAL NETWORKS

This is a feasibility study on training an Artificial Neural Network ( ANN) classifier to detect mass regions on mammograms, using a database consisting of 87 clinical mammograms. Texture features extracted from manually selected regions of interest in the mammograms, including ma sses and normal breast parenchyma, were input into a three-layer feed- forward ANN. The data were divided into five groups, and different com binations of these groups formed four sets of training and test data. We achieved on the average a true positive fraction of 84% at a false positive fraction of 34% with an ambiguity rate of 5%. We did not obse rve performance improvement with a four-layer ANN. This pilot study pa ves the way for further studies in classification of different types o f masses and normal breast parenchyma.