Generating ROC curves for artificial neural networks

Receiver operating characteristic (ROC) analysis is an established method of measuring diagnostic performance in medical imaging studies. An ROC curve characterizes the inherent tradeoff between true positive and false positive detection rates in a classification system. Traditionally, artificial neural networks (ANNs) have been applied as a classifier to find one "best" partition of feature space, and therefore a single detection rate. This work proposes and evaluates a new technique for generating an ROC curve for a 2-class ANN classifier. We show that the new technique generates significantly better ROC curves than the method currently used to generate ROCs for ANNs.<>