Evaluating detector resolution compensation methods in SPECT imaging through numerical observer ROC and human observer LROC

The channelized Hotelling observer (CHO) has been shown to correlate with humans for a variety of detection tasks. In a previous investigation, the authors found that the average performance over tumor location of a CHO in a “signal-known-exactly” ROC study could be used to eliminate some suboptimal imaging strategies from consideration in a human localization ROC (LROC) study. Here, that conclusion is tested for a human LROC study intended to evaluate detector resolution compensation (DRC) in OSEM reconstruction. The detection task was “hot” tumor detection in Ga-67 scans of the chest. Images were obtained with 1, 2, and 4 iterations of OSEM with attenuation correction (AC) and Gaussian diffusion DRC, and with 1 iteration of OSEM with just AC. Human results were based on 120 images (60 tumor locations) for each of the strategies, while the CHO was applied to 100 images at each of 5 tumor locations. The rankings of the strategies by the CHO and the average human observer showed agreement when similar nonlinearly processed images were used by both observers. Poorer correlation was found when the CHO used only linearly processed images. This raises questions about whether a recent theoretical formulation of the noise properties of iterative algorithms can be used to form the CHO for the authors´ LROC studies. This will be investigated in upcoming studies