ABSTRACT
This chapter describes various empirical operating characteristics that can be plotted from FROC data. The chapter is divided into two parts: formalisms and examples. A distinction is made between latent (suspicious regions, regardless of whether they are marked) and actual marks. The formalism describes the construction of operating points from the data. Calculation of corresponding AUCs is addressed in the next chapter. Formalism for the AFROC, the inferred ROC, the weighted AFROC (wAFROC), etc., is presented. R-code examples of raw and binned FROC, AFROC plots are shown. A misconception about location-level true negatives – actually an unmeasurable event – is explained. The rationale is presented for not using non-lesion localizations on diseased cases in computing a measure of performance. Another misconception, that neglect of all but the highest rated NL on non-diseased cases sacrifices much statistical power, is addressed. It is recommended that FROC curve-based measures of performance not be used to assess performance in localization tasks. Evidence for this recommendation, based on simulation studies, is presented here and in Chapter 17. The AUC under the weighted AFROC is the recommended way of summarizing performance in localization tasks.
