Signal detection in lumpy backgrounds

Abstract

In this dissertation we explore signal detection with model and human observers in the setting of nuclear medicine. Regarding the model observer, the main focus is on the linear observer that maximizes detectability, which we call the Hotelling observer. In particular, we outline two methods for realizing an estimate of this observer. The first uses a Fourier representation. The second uses a representation with a small number of channels chosen for optimal performance. The work employs statistically defined lumpy backgrounds to test the methods and results. These backgrounds are more complicated than correlated Gaussian noise and are meant to complicate the signal-detection task by involving random structure. Regarding the human observer, we present a literature review of psychophysical models, including results based on these models. We then examine one current front runner--a channelized-Hotelling observer with channels modeling visual-response functions---for two experiments involving the lumpy backgrounds.