| The quantitative assessment of medical image quality should be tied to specific tasks. The use of receiver operating characteristic (ROC) studies to analyze human observer performance is considered the ‘gold standard’ for detection task studies. These studies, however, require considerable time and effort and so are not well-suited for parameter optimization. Thus there is considerable interest in developing mathematical ‘model’ observers that mimic the performance of human observers.; The detection task is an essential component of Positron Emission Tomography (PET) imaging, which is routinely used for the detection and staging of cancer. PET imaging systems acquire and reconstruct tomographic images as contiguous volumetric (3D) images. Physicians typically interpret these images by searching the image volume using volumetric displays (showing linked transverse, coronal and sagittal sections of the image volume). Most observer studies, however, typically use planar images for evaluation with both human and model observers.; This dissertation proposes a volumetric model observer computation method, and investigates differences between planar (2D) and volumetric (3D) model observer performances and their correlation with human observer performance for PET imaging. Multiple sinogram realizations were simulated with statistically accurate noise properties. Sinograms were corrected for attenuation and background coincidences, and reconstructed with filtered backprojection (FBP) into image volumes. The apodizing filter used for FBP was the Hanning window. For human observer ROC studies, an observer scoring software tool, similar to the volumetric display being used in clinical PET imaging, was developed. For model observer studies, the non-prewhitening matched filter (NPWMF) and the channelized Hotelling observer (CHO) were used, in both 2D and 3D implementations.; The 3D NPWMF and the CHO model observers yielded significantly higher detectabilities than the 2D model observers, and had higher correlations with human performance than the 2D versions. When the axial smoothing was applied to achieve isotropic target resolution, however, performance differences between the 2D and the 3D model observers were not significant.; Because of their high correlations with human performance, the volumetric model observers, especially the 3D CHO, can be used as surrogates for measuring human performance, in particular for optimizing free parameters for cancer detection with whole-body PET imaging. |
