Model observers to predict human performance in LROC studies of SPECT reconstruction using anatomical priors [6917-26]

We investigate the use of linear model observers to predict human performance in a localization ROC (LROC) study. The task is to locate gallium-avid tumors in simulated SPECT images of a digital phantom. Our study is intended to find the optimal strength of smoothing priors incorporating various degrees of anatomical knowledge. Although humans reading the images must perform a search task, our models ignore search by assuming the lesion location is known. We use area under the model ROC curve to predict human area under the LROC curve. We used three models, the non-prewhitening matched filter (NPWMF), the channelized nonprewhitening (CNPW), and the channelized Hotelling observer (CHO). All models have access to noise-free reconstructions, which are used to compute the signal template. The NPWMF model does a poor job of predicting human performance. The CNPW and CHO model do a somewhat better job, but still do not qualitatively capture the human results. None of the models accurately predicts the smoothing strength which maximizes human performance.