MetadataShow full item record
PublisherThe University of Arizona.
RightsCopyright © is held by the author. Digital access to this material is made possible by the University Libraries, University of Arizona. Further transmission, reproduction or presentation (such as public display or performance) of protected items is prohibited except with permission of the author.
AbstractAn important goal in medical imaging is the assessment of image quality in a way that relates to clinical efficacy. An objective approach is to evaluate the performance of diagnosis for specific tasks, using ROC analysis. We shall concentrate here on classification tasks. While many factors may confine the performance achieved for these tasks, we shall investigate two main limiting factors: image blurring and object variability. Psychophysical studies followed by ROC analysis are widely used for system assessment, but it is of great practical interest to be able to predict the outcome of psychophysical studies, especially for system design and optimization. The ideal observer is often chosen as a standard of comparison for the human observer since, at least for simple tasks, its performance can be readily calculated using statistical decision theory. We already know, however, of cases reported in the literature where the human observer performs far below ideal, and one of the purposes of this dissertation is to determine whether there are other practical circumstances where human and ideal performances diverge. Moreover, when the complexity of the task increases, the ideal observer becomes quickly intractable, and other observers such as the Hotelling and the nonprewhitening (npw) ideal observers may be considered instead. A practical problem where our intuition tells us that the ideal observer may fail to predict human performance occurs with imaging devices that are characterized by a PSF having long spatial tails. The investigation of the impact of long-tailed PSFs on detection is of great interest since they are commonly encountered in medical imaging and even more generally in image science. We shall show that the ideal observer is a poor predictor of human performance for a simple two-hypothesis detection task and that linear filtering of the images does indeed help the human observer. Another practical problem of considerable interest is the effect of background nonuniformity on detectability since, it is one more step towards assessing image quality for real clinical images. When the background is known exactly (BKE), the Hotelling and the npw ideal observers predict that detection is optimal for an infinite aperture; a spatially varying background (SVB) results in an optimum aperture size. Moreover, given a fixed aperture size and a BKE, an increase in exposure time is highly beneficial for both observers. For SVB, on the other hand, the Hotelling observer benefits from an increases in exposure time, while the npw ideal observer quickly saturates. In terms of human performance, results show a good agreement with the Hotelling-observer predictions, while the performance disagrees strongly with the npw ideal observer.
Degree ProgramOptical Sciences