Performance evaluation of the channelized Hotelling observer using bootstrap list-mode PET studies

Abstract

This study investigated whether list-mode PET data generated using the bootstrap method can be used to predict lesion detectability as assessed by the channelized Hotelling observer (CHO). A Monte-Carlo simulator was used to generate 2D PET list-mode data set acquisitions of a disk object. One of these list-mode sets was then used to create an ensemble of bootstrap list-mode sets. A randomly positioned signal (lesion) was introduced into half of the list-mode sets to create an ensemble of signal-present and signal-absent list-mode sets. These sets were then reconstructed using the OSEM list-mode algorithm. The CHO was computed from the ensemble of reconstructed images generated from the bootstrap data sets as well as from independent noisy data sets. The F-test and the student t-test found no significant difference (confidence level 5%) in the areas under the LROC curve generated using the independent noisy list-mode sets and the bootstrap list-mode sets for clinical count levels. It is also shown how bootstrap images can be used to implement a patient-specific, CHO-based stopping-rule criterion for ordered-subset expectation-maximization (OSEM) list-mode iterative reconstruction. An example of applying the CHO-based stopping-rule criterion for list-mode reconstruction of the MCAT phantom showed an optimal detectability index at iterations 7 using 2 subsets respectively. Results from this study suggest that the bootstrap approach can be used to conduct numerical observer studies with more realistic backgrounds by generating them from a patient study (with the introduction of simulated lesions), and allows the possibility of applying a patient-specific, CHO-based stopping-rule criterion for list-mode iterative reconstruction.