The FROC, AFROC and DROC Variants of the ROC Analysis

Receiver operating characteristic (ROC) methodology [1-€“5] is widely used in diagnostic radiology to measure imaging task performance. A ROC experiment yields a measure Az, the area under the ROC curve, which is independent of disease prevalence and the criteria used by the observer in reporting decisions. The ROC method is strictly applicable only to tasks that call for a binary decision on the part of the observer: is the image normal or abnormal, and most diagnostic tasks do not satisfy this requirement. For example, in nodule detection in chest radiography, the clinical task requires specification of location(s) of the perceived nodule(s). This information is of clinical relevance as it may guide subsequent surgical intervention, yet this information must be ignored in ROC studies, resulting in an imprecise or fuzzy scoring method. In particular, if the observer's decision was abnormal for an actually positive image but he indicated an incorrect location, the ROC paradigm would not penalize this reader. This event would be scored as a true positive even though the reader committed two errors on this image: missing the actual lesion (local false negative) and selecting a location that was nodule free (local false positive). Another issue is that when multiple lesions may present on the same image, ROC analysis ignores additional information (more than 1 perceived lesion) that the reader is prepared to provide. For an image with two nodules, one reader may detect only one of the nodules while the other detects both nodules. Both readings would be scored identically in the ROC method. However, the clinical consequences of missing one of the nodules could be very significant.