Performance of model observers for lesion detection in SPECT imaging

Abstract

We investigated the performance of model observers (the channelized Hotelling observer (CHO), channelized non-prewhitening observer (CNPWO) and non-prewhitening observer (NPWO)) for lesion detection in single photon emission computed tomography (SPECT) imaging. Using a simulation tool, we generated the numerical phantom with a hot lesion for cases having various contrasts and sizes and projection data including photon attenuation, scatter and distance-dependent collimator-detector response. The SPECT images were reconstructed using the filtered back-projection method and the ordered-subsets expectation-maximization method. Fifty lesion-absent and lesion-present images were generated and the detectability indices derived from CHO (d(CHO)), CNPWO (d(CNPWO)) and NPWO (d(NPWO)) were calculated from a total of 100 images. These procedures were repeated 5 times for each condition. The relationships among these detectability indices were investigated for various projection counts, lesion contrasts and lesion sizes. All detectability indices increased with increases of projection count, lesion contrast and lesion size. Although there were good correlations among them, the relationships between d(CHO) and d(NPWO) and between d(CNPWO) and d(NPWO) were not linear. The coefficient of variation of d(CHO) tended to be significantly greater than that of d(CNPWO) and d(NPWO). This study will be useful for understanding the performance of model observers when evaluating lesion detection in SPECT imaging.