Nonparametric estimation of degenerate ROC data sets used for comparison of imaging systems.

Maximum likelihood estimation, assuming a binormal model, has been used extensively to estimate the area under receiver operating characteristic (ROC) curves and thus provide an index of diagnostic accuracy for the comparison of two imaging systems. However, in some instances, a degenerate data set results in a fitted ROC curve of inappropriate shape. In those instances, a nonparametric Wilcoxon statistic may be used to compare the areas under two ROC curves. Simulation of this approach indicates that the procedure has relatively high statistical power and its use in specific degenerate data sets shows its applications potential.