Analysis of observer performance in detecting signals with location uncertainty for regularized tomographic image reconstruction

Our goal is to optimize regularized image reconstruction methods for emission tomography with respect to the task of detecting small lesions in the reconstructed images. To reflect medical practice realistically, we consider the location of the lesion to be unknown. This location uncertainty significantly complicates the mathematical analysis of model observer performance. We consider model observers whose decisions are based on finding the maximum value of a local test statistic over all possible locations. Khurd and Gindi (SPIE 2004) and Qi and Huesman (SPIE 2004) described analytical approximations of the moments of the local test statistics and used Monte Carlo simulations to evaluate the localization performance of such "maximum observers". We propose here an alternative approach, where tail probability approximations developed by Adler (AAP 2000) facilitate analytical evaluation of the detection performance of these observers. We illustrate how these approximations can be used to evaluate the probability of detection (for low probability of false alarm operating points) for the maximum channelized hotelling observer. Using our analyses, one can rank and optimize image reconstruction methods without requiring time-consuming Monte Carlo simulations.