PURPOSE: To evaluate the performance and inter- and intraobserver vari
ability of an artificial neural network (ANN) for predicting breast bi
opsy outcome. MATERIALS AND METHODS: Five radiologists described 60 ma
mmographically detected lesions with the American College of Radiology
Breast Imaging Reporting and Data System (BI-RADS) nomenclature. A pr
eviously programmed ANN used the BI-RADS descriptors and patient histo
ries to predict biopsy results. ANN predictive performance was compare
d with the clinical decision to perform biopsy. Inter- and intraobserv
er variability of radiologists' interpretations and ANN predictions we
re evaluated with Cohen kappa analysis. RESULTS: The ANN maintained 10
0% sensitivity (23 of 23 cancers) while improving the positive predict
ive value of biopsy results from 38% (23 of 60 lesions) to between 58%
(23 of 40 lesions) and 66% (23 of 35 lesions; P < .001). Interobserve
r variability for interpretation of the lesions was significantly redu
ced by the ANN (P < .001); there was no statistically significant effe
ct on nearly perfect intraobserver reproducibility. CONCLUSION: Use of
an ANN with radiologists' descriptions of abnormal findings may impro
ve interpretation of mammographic abnormalities.