COMPUTER-AIDED DIAGNOSIS OF BREAST-CANCER - ARTIFICIAL NEURAL-NETWORKAPPROACH FOR OPTIMIZED MERGING OF MAMMOGRAPHIC FEATURES

Rationale and Objectives. An artificial neural network (ANN) approach was developed for the computer-aided diagnosis of mammography using an optimally minimized number of input features. Methods. A backpropagat ion ANN merged nine input features (age plus eight radiographic findin gs extracted by radiologists) ro predict biopsy outcome as its output. The features were ranked, and more important ones were selected to pr oduce an optimal subset of features. Results. Given all nine features, the ANN performed with a receiver operator characteristic area under the curve (A(z)) of .95 +/- .01. Given only the four most important fe atures, the ANN performed with an A(z) of .96 +/- .01. Although nor si gnificantly better than the ANN with all nine features, the ANN with t he four optimized features was significantly better than expert radiol ogists' A(z) of .90 +/- .02 (p = .01). This four-feature ANN had a 95% sensitivity and an 81% specificity. For cases with calcifications, th e radiologists' performance dropped to an A(z) of .85 +/- .04, whereas a specialized three-feature ANN performed significantly better with a n A(z) of .95 +/- .02 (p = .02). Conclusion. Given only four input fea tures, the ANN predicted biopsy outcome significantly better than did expert radiologists, who also had access to other radiographic and non radiographic data. The reduced number of features would substantially decrease data entry efforts and potentially improve the ANN's general applicability.