Evolving artificial neural networks for screening features from mammograms.

摘要: Abstract Disagreement or inconsistencies in mammographic interpretation motivates utilizing computerized pattern recognition algorithms to aid the assessment of radiographic features. We have studied potential for using artificial neural networks (ANNs) analyze interpreted features from film screen mammograms. Attention was given 216 cases (mammogram series) that presented suspicious characteristics. The domain expert (Wasson) quantified up 12 each case based on guidelines previous literature. Patient age also included. existence absence malignancy confirmed via open surgical biopsy (111 malignant, 105 benign). ANNs various complexity were trained evolutionary programming indicate whether not a present vector scored input statistical cross validation procedure. For masses, best evolved generated mean area under receiver operating characteristic curve ( A Z ) 0.9196±0.0040 (1 S.E.), with specificity 0.6269±0.0272 at 0.95 sensitivity. Results when microcalcifications included quite as good =0.8464), however, only two hidden nodes performed well more complex and better than one node. performance comparable prior literature, but an order magnitude less complexity. success small diagnosing breast cancer offers promise suitable explanations ANN's behavior can be induced, leading greater acceptance by physicians.