<title>Free-response receiver operating characteristic analysis in medical imaging</title>

Abstract

Recently we have introduced a method for analyzing Free-response Receiver Operating Characteristic (FROC) experiments. These differ from traditional ROC observer performance experiments by allowing for multiple abnormalities and observer responses per image, and by requiring correct localization for true positive events. We introduced the Alternative FROC (AFROC) plot which is analogous to the ROC plot, and the area under it, which is analogous to the area parameter of ROC methodology. Subsequently we have extended the analysis with a new way of scoring FROC images, termed AFROC analysis. This renders unnecessary the Poisson assumption used in the earlier paper. We also introduced the Free-response Forced Error (FFE) experiment which renders unnecessary the Gaussian distribution assumption of FROC and AFROC analysis. FROC/FFE experiments performed with synthetic images (disk objects at random locations on a noise background) which illustrate the FROC methodology are presented. We wish to bring to the attention of PACS researchers that rather sophisticated efficacy analysis, which has several advantages over the standard ROC approach, is made possible through this methodology.