Effects of noise, image resolution, and ROI definition on the accuracy of standard uptake values: A simulation study

Semiquantitative standard uptake values (SUVs) are used for tumor diagnosis and response monitoring. However, the accuracy of the SUV and the accuracy of relative change during treatment are not well documented. Therefore, an experimental and simulation study was performed to determine the effects of noise, image resolution, and region-of-interest (ROI) definition on the accuracy of SUVs. Methods: Experiments and simulations are based on thorax phantoms with tumors of 10-, 15-, 20-, and 30-mm diameter and background ratios (TBRs) of 2, 4, and 8. For the simulation study, sinograms were generated by forward projection of the phantoms. For each phantom, 50 sinograms were generated at 3 noise levels. All sinograms were reconstructed using ordered-subset expectation maximization (OSEM) with 2 iterations and 16 subsets, with or without a 6-mm gaussian filter. For each tumor, the maximum pixel value and the average of a 50%, a 70%, and an adaptive isocontour threshold ROI were derived as well as with an ROI of 15 x 15 mm. The accuracy of SUVs was assessed using the average of 50 ROI values. Treatment response was simulated by varying the tumor size or the TBR. Results: For all situations, a strong correlation was found between maximum and isocontour-based ROI values resulting in similar dependencies on image resolution and noise of all studied SUV measures. A strong variation with tumor size of ≥50% was found for all SUV values. For non-smoothed data with high noise levels this variation was primarily due to noise, whereas for smoothed data with low noise levels partial-volume effects were most important. In general, SUVs showed under- and overestimations of ≥50% and depended on all parameters studied. However, SUV ratios, used for response monitoring, were only slightly dependent of ROI definition but were still affected by noise and resolution. Conclusion: The poor accuracy of the SUV under various conditions may hamper its use for diagnosis, especially in multicenter trials. SUV ratios used to measure response to treatment, however, are less dependent on noise, image resolution, and ROI definition. Therefore, the SUV might be more suitable for response-monitoring purposes.