senior radiologists.27 In another study, use of a four-point rating scale in comparing tests of early pregnancy yielded the same accuracy index for days of amenorrhea, uterine size, and the ratio of the two; and a somewhat greater accuracy for ultrasound.8 In a study of the quantitative characteristics of normal and suspicious cytologic smears, investigators used an equivalent of the ROC to evaluate screening performance for smears collected by vaginal aspiration, cervical swab, and cervical-scraping techniques,56 and found substantial differences in accuracy between the three techniques.
The last study to be mentioned examined four radiographic screen-film systems, found good reproducibility from one observer to another, and showed average values of accuracy ranging from rather low () to very high ( ). The investigators point out that a combination of film and screen yielding is about twice as fast as the combinations yielding values of of 1.3 and 3.1, resulting in about half the patient exposure,17,30 and four times faster than the system yielding .* In this setting, then, a good basis exists for trading accuracy and hazard.
It may be helpful to point out that the relatively well established imaging techniques used in studies of reader reliability and reader experience (mentioned early in this section), yielded values of Δm, d́, orin the vicinity of 2.5. Those studies which focused on newer physical imaging techniques (with the exception of the last-mentioned studies, which covered nearly the full range of accuracy), yielded values of these accuracy indices near 1.5. One can refer to Fig. 2 to get an appreciation of the error rates associated with indices of 2.5 and 1.5. Taking points along the minor diagonal, where the probabilities of false-negatives and false-positives are equal, the probability of each error is about 0.10 for the index of 2.5 and in excess of 0.20 for the index of 1.5.
An ROC measure of accuracy is appropriately used in evaluations of imaging techniques and image readers, because analysis and experiments have indicated that other available measures will confound the observer's decision bias with the variable of principal interest: either the diagnostic accuracy afforded by the imaging technique, or the perceptual capability of the observer. Several studies have demonstrated the practicality of the medical-data gathering procedure associated with the ROC. Other studies also suggest that the ROC's basis in decision theory helps to establish a link between the various decision parameters the observer can reflect and the medical and social utility of a diagnostic system.28
Several questions of theory and method involving the medical application of the ROC analysis have arisen earlier in the ROC's application to human sensory____________________