Assessment of PET and SPECT phantom image quality through automated binary classification of cold rod arrays

Purpose Evaluation of positron emission tomography (PET) or single photon emission computed tomography (SPECT) performance often involves qualitative assessment of phantom images, namely the visibility of hot or cold structures in a warm background. Structure‐to‐background contrast is a quantitative measure of scanner performance; however, contrast measurements do not account for image noise and its effect on task performance. Although task‐based performance could be evaluated over an ensemble of phantom scans, a more practical approach is desired. Methods Repeated structures, such as the cold rod arrays of the American College of Radiography (ACR) PET and SPECT phantoms, offer an opportunity to evaluate image quality based on statistical decision theory. Images are co‐registered to a digital template for placement of numerous regions of interest (ROIs) in multiple thick slices centered on each cold rod and on each midpoint between rods. The assumption is made that each ROI corresponds to a statistically independent measurement, known to be of either a cold rod or the background. A receiver operating characteristic (ROC) curve then is generated for each cold rod sector of the phantom. The area under the ROC curve (AUC) provides a quantitative measure of cold rod visibility. In addition, signal‐to‐noise ratio (SNR) is calculated, under the assumption that the rod and background ROI measurements are distributed normally. Results Using this approach, data from PET and SPECT phantom studies from prior annual physics surveys were analyzed retrospectively. Rod and background ROI measurements had nearly normal (PET) or approximately normal (SPECT) distributions. Resultant ROC curves illustrated the varying degrees of overlap between rod and background histograms vs cold rod diameter. Both AUC and SNR correlated well with visual image assessment and had high consistency between phantom studies, demonstrating the quantitative accuracy of the automated analysis. AUC reliably quantified cases where cold rods were partially or mostly resolved, while SNR provided further characterization when cold rods were completely resolved. In comparison, contrast measurements often mirrored AUC and SNR but were inconsistent in cases of varying noise. Conclusions Automated analysis of phantom cold rod arrays using binary classification provides informative quantitative measures of image quality and is practical for routine use. This approach avoids observer dependencies associated