Predictions of the RSM

This chapter describes predictions of the RSM and how they compare with evidence. All RSM-predicted operating characteristics share a constrained end-point property. All other ROC models predict that the end-point, namely the uppermost non-trivial point on the ROC, reached at infinitely low reporting threshold, is (1,1), while the RSM predicts it does not reach (1,1). The reason is that the RSM allows the existence of cases with no latent sites and hence no Z-samples. Expressions for the ROC operating point as a function of the three RSM parameters, the lesion distribution vector, and the reporting threshold, are derived. The accessible portion of the RSM-predicted ROC curve is proper. To fully account for performance, the area under a straight-line extension from the empirical end-point to (1,1) needs to be included. Similar comments apply to the AFROC and wAFROC. Unlike them, the FROC curve cannot be meaningfully extended beyond the observed end-point. Chance level performance for the AFROC and the observer who does not mark any image is discussed. The latter, yielding AFROC-AUC = 0.5, is more informative - perfect performance on non-diseased cases - compared to the corresponding ROC (ROC-AUC = 0.5). Expressions are derived for search and lesion-classification performance, defined in Chapter 15, as functions of RSM parameters. These can be estimated using RSM-based curve fitting described in the next chapter. Evidence for the validity of the RSM is presented as is further evidence that the FROC is a poor descriptor of performance.