ABSTRACT
Because of degeneracy issues, where two RSM parameters appear in an inseparable combination, FROC curve-based fitting is not possible with human observer data. However, it can be used for designer-level CAD data, as in the initial-detection candidate-analysis (IDCA) method. Fitting the ROC curve, which breaks the degeneracy, has proven successful. The RSM-ROC fitting algorithm was applied, along with PROPROC and CBM, to 236 individual datasets, drawn from 14 MRMC datasets described in Online Chapter 24. A sampling of the three fits are presented in this chapter, and the rest can be viewed on the author's website. For each dataset, AUCs predicted by PROPROC or CBM, when plotted against RSM AUCs, showed a near unit-slope linearity through the origin, with R2 > 0.999. Compared to the RSM, the PROPROC predictions were, on the average, 2.6% larger, while the CBM estimates were 1% larger. In some cases, PROPROC, and to a lesser extent, CBM, grossly overestimated AUC by performing a large extrapolation to (1,1) unsupported by the intermediate operating points. The near equality of all three estimates is explained in terms of uniqueness of ideal observer performance that each algorithm attempts to emulate. An inverse correlation was found between search and lesion-classification performances, suggesting that observers tend to compensate for deficiency in search by improved performance in lesion-classification, and vice-versa. Average search performance was 23%, average lesion-classification performance was 88%, and average AUC was 79%. These values demonstrate that search expertise is the bottleneck limiting overall performance in diagnostic tasks.
