Application of artificial neural networks in mammography for the diagnosis of breast cancer
Chris Y. Wu, Maryellen Lissak Giger, Kunio Doi, Charles E. Metz, Robert M. Nishikawa, Carl J. Vyborny, Robert A. Schmidt · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1992
Yuzheng Wu, Maryellen L. Giger, Kunio Doi, Charles B. Metz,Robert M. Nishikawa, Carl J. Vyborny, Robert A. SchmidtKurt Rossmann Laboratories for Racliologic Image ResearchDepartment of Radiology, MC2026The University of ChicagoChicago, illinois 60637ABSTRACTInterpretation of mammograms for the diagnosis of breast cancer is a difficult task. Weinvestigated the potential utility of neural networks in the analysis of mammographic data as an aidto radiologists in arriving at correct diagnoses. Three-layer, feed-forward networks with a back-propagation algorithm were employed in this study. The neural networks were trained for theinterpretation of mammograms based on: (a) features extracted from mammograms by expertradiologists; and (b) radiographic patterns of clustered microcalcifications in digital mammograms,as represented by pixel values.Performance of the neural networks was evaluated by ROC analysis with jack-knifemethod on individual cases. The network that used human-extracted features as input performedwell in distinguishing between benign and malignant lesions, yielding an A value of 0.95 in a testby the round-robin method for textbook cases. The performance of a neural network in mergingradiologist-extracted features of lesions to diagnose breast cancer was found to be higher than theaverage performance of attending and resident radiologists alone (without the aid of a neuralnetwork). The networks trained on digital mammograms also performed at a fairly high level,achieving an A value of 0.88 in detecting microcalcffications and demonstrated the potential tohelp reducing false-positive clusters produced by the rule-based detection scheme ofmicrocalcifications developed in our lab. The neural networks are capable of classifyingmammographic images on the basis of both extracted features and radiographic patterns, andtherefore, are potentially useful tools for mammographic interpretation.1. INTRODUCTIONArtificial neural networks, which constitute a non-algorithmic approach to informationprocessing, have been studied intensively in the field of computer science in recent years.13 Theseneural networks, which are capable of parallel-processing a large amount of informationsimultaneously, address problems not by pre-specified conventional algorithms, but rather bylearning from examples presented repeatedly. The popularity of neural networks is primarily dueto their apparent ability to make decisions and draw conclusions when presented with complex,noisy, or partial information and to adapt their behavior to the nature of the training data.Neural networks have been applied to medical imaging and decision-making in recent yearsand have been shown to be a powerful tool for pattern recognition and data classification.414Among the applications, neural networks have been employed in attempts to interpret neonatal