Accounting for anatomical noise in tumor localization tasks with visual-search and channelized Hotelling observers

Several scanning observers have been proposed for tumor detection-localization tasks in SPECT. Performance of these tasks is complicated by anatomical noise present in the images. We compare the scanning chanellized Hotelling observer (CHO) and the visual-search (VS) observer as two ways of accounting for anatomical noise in these tasks. SPECT projections are generated with an analytic projector. The ordered subset expectation maximization algorithm is used to reconstruct the projections with ten different reconstruction strategies. Localization ROC studies are conducted with the scanning CHO, two forms of the VS observer, the scanning channelized nonprewhitening observer and a set of human observers. The correlations between human and model observers are analyzed. Results show that the VS observers have a higher correlation with the human observers as compared to the scanning observers.