This research will study and compare observers' performance in a variety of prototypical detection and feature-discrimination tasks, using high-quality digital images with uncorrelated-noise backgrounds. The estimates of human observers' detection and discrimination capabilities will also be compared with the ideal observer's capability to perform the same decision task with the same physical images. This will permit estimates of observers' efficiency, relative to the maximum possible performance. Our previous research with CT images indicates that changes in observers' ability to detect lesions on those images can be predicted by the performance of a physical cross-correlator, across a wide variety of manipulations of the physical image properties. This research will extend our previous investigations to images in which the cross-correlator is an optimum physical calculation, and to tasks that are more complex than feature detection. The measurements of observer efficiency will be interpreted by a model of the human observer, which attributes the observer's reduced efficiency to a combination of systematic factors and unsystematic sources of variability (observer noise). This model of the observer will be developed and tested in detection and discrimination tasks, whose conditions are deliberately designed: (a) to manipulate potential systematic and unsystematic sources of observer inefficiencies, and (b) to alter the relative importance of the physical-noise and observer-noise in limiting the observer's ability to perform the decision task. The proposed research will also develop and test new ROC methods for evaluating a combination of the observer's likelihood ratings and identification judgments in task that present more than two specified alternatives.
Showing the most recent 10 out of 12 publications
