Abstract
In the past fifty years, a large number of ROC curves have been generated by radiologists reading medical images. There is a common belief that human ROC curves should generally be convex and that non-convex ROC curves are caused by some imperfections in the measurement procedures, for example, the limited number of images read by observers, poor experimental design and/or curve fitting errors. We propose an ideal observer (IO) model to describe a radiologist’s performance based on the fact that a convex ROC curve corresponds to one unique pair of likelihood ratio (LR) distributions. We define an equivalent IO (EIO) as the one who has the same performance as the human observer, as characterized by an ROC curve. To measure the LR distributions of the EIO, we formalize experimental design principles that force the observers to act rationally based on von Neumann and Morgenstern’s axioms. We show that human observer study design refinements in radiology, although motivated by empirical or practical principles, implicitly enforce rationality and result in reasonably convex ROC curves. EIO theory allows us to model the human observer as an ideal observer. It also suggests that while the human observers are not optimal or cannot fully use the statistical information in medical images, at some level they can be rational or ideal in using the limited information that they are able to extract from the images for decision making.
Meeting abstract presented at VSS 2013