Comparison of hotelling observer models and human observers in defect detection from myocardial SPECT imaging

The objective of this study was to compare defect detection performance usi ng a channelized Hotelling observer with radially symmetric channel (RSC) a nd oriented channels (OC) to that found in a previously reported human obse rver ROC study [1, 2]. The observer's task involved the detection of left-v entricular myocardial perfusion defects for Tc-99m sestamibi myocardial per fusion SPECT images. The images were reconstructed using the filtered-backp rojection (FBP) algorithm without attenuation compensation or the maximum-l ikelihood expectation-maximization (ML-EM) algorithm with non-uniform atten uation compensation (AC), This was performed for cases with and without sig nificant attenuation artifacts. The areas under the ROC curves for the Hote lling observers were calculated and compared to those for the human observe rs. The ML-EM reconstructed images (with AC) had high defect detectability across all anatomy types, while with FBP, a lower detectability was found f or cases where the reconstructed images contained attenuation artifacts in the myocardium. Similar trends between the channelized Hotelling observer a nd those from the human observer study were found with both radially symmet ric channel (RSC) and oriented-channel (OC) models. Further investigation o f the channel models is needed to determine the number and alignment of ori entations and the number and cutoff of frequency bands to improve agreement between human and Hotelling observer results.