Model observers for predicting human performance on signal detection tasks.

Abstract

Various model observers have been applied to the objective assessment of medical image quality. However, the relevance of this application to clinical efficacy depends largely on how well model observers predict human performance. Attempting to answer this question, this dissertation focuses on the investigation of a linear observer known as the Hotelling observer and a modified version of the Hotelling observer, known as the channelized Hotelling observer. Performances of these observers for a signal-known-exactly detection task are calculated and compared to the performance of the human observer. Several psychophysical studies suggest that the Hotelling observer, formulated on the first- and second-order statistical properties of the images, could predict the human performance very well. To investigate the effect of certain higher-order statistical information on human performance, an experiment was designed in which the mean, variance, and covariance of three groups of images were kept the same, while the shapes of the image grey-level histogram were varied. The results showed little practical difference in the human performance among the three groups; thus the higher-order statistical information represented by the shape of the grey-level histogram did not influence the human observer's signal-detection performance for the task considered in this experiment. Another linear model observer, the nonprewhitening observer has been found in previous work to predict human performance better than the Hotelling observer for images with uniform backgrounds and correlated noise. When the images contain nonuniform background and uncorrelated noise, however, the Hotelling observer is found to be better in predicting human performance. To unify these results, a channelized Hotelling observer was proposed whose performance resembles that of a nonprewhitening observer for images with correlated noise, and that of a Hotelling observer for images with nonuniform background. Moreover, the channelized Hotelling observer is able to predict human performance when images have both the nonuniform background and correlated noise. A nonlinear version of the channelized Hotelling observer has also been found to predict human performance well.