ABSTRACT
This chapter introduces the FROC paradigm. Unlike the ROC paradigm, this paradigm allows for localization information, classified as correct or incorrect, to be used in the analysis. The FROC paradigm is shown to be a search task. The structure of FROC data, namely a random number of zero or more mark-rating pairs per image, is described. A proximity criterion is used to classify each mark as lesion localization or non-lesion localization. The use of ROC terminology such as true positive and false positive in the FROC context is discouraged. The FROC paradigm is placed in its historical context and a key publication by Bunch et al. is described. The FROC curve is introduced and a data simulator, implemented in R, is used to show its dependence on perceptual signal-to-noise ratio (pSNR). A solar analogy is used to explain, at an intuitive level, the dependence of the FROC curve on pSNR. Online appendices explain the code used to generate the plots, physical SNR measurements in the mammography quality control context, and the Bunch transforms relating ROC and FROC curves.
