Abstract
372
Objectives PET SUV measures are inaccurate due to bias from finite resolution and variance from quantum noise. Reconstruction algorithms that control the tradeoff between bias and variance vary between sites and manufacturers and confound comparing intra- or inter-site PET images. We present a process to characterize the agreement between algorithms that incorporates both bias and variability as a post-hoc method to harmonize PET reconstruction protocols.
Methods A long-lived Ge-68 phantom based on the NEMA NU-2 Image Quality phantom was constructed with a 7.7:1 feature:background ratio. We varied emission counts, reconstruction methods, and filtering and acquired 25 emission data sets per parameter set on different commercial PET/CT scanners. The normalized root-mean-square difference (NRMSD) is defined as the RMS difference between pairs of recovery coefficient curves. The NRMSD was calculated over the 25x25 realizations for each combination of parameters for each pair of scanners.
Results Over the range of parameters tested, the NRMSD between pairs of recovery coefficient curves (for SUVmax) reached a maximum of 29% with a median value of 13%, reflecting the potential range of quantitative errors in inter-scanner comparisons. For clinically relevant acquisition and reconstruction parameters, the best-case pair of parameter sets for two different PET scanner manufacturers led to a NRMSD (for SUVmax) of only 4%, with a maximum per-sphere difference of 6%.
Conclusions The use of a long-lived phantom for repeated acquisitions, combined with the normalized root-mean-square difference (NRMSD) metric, allows determining which sets of acquisition and reconstruction parameters lead to the lowest difference in image quantitation. Matching parameters in this post-hoc manner allows the creation of protocols that provide SUV metrics with minimized bias and variability between image sets. Further work is needed to include the absolute bias effects and account for task dependence.
Research Support Supported by NIH grant R01 CA169072 and SAIC Contract 24XS036-004