Influence of three reconstruction algorithms on the estimation of the standardize uptake value in ¹⁸F-fluoride PET

In clinical practice accurate and precise quantification of PET images is of crucial importance especially for tumor staging when comparing pre- and post- treatment scans. For example a change in the standardized uptake values (SUV) of 10-20% can discriminate whether a treatment is effective or not. However, several factors cause degradation of the accuracy and precision of quantitative information. Iterative reconstruction methods can help towards more accurate tumor delineation and potentially quantification. However, the most commonly used reconstruction technique (Ordered Subset Expectation Maximization, OSEM) has pitfalls that are critical for quantification: bias (e.g. at early iterations and at low count statistics) and noise (e.g. at late iterations). The aim of this work is to evaluate the quantitative performance of existing reconstruction algorithms with acquired phantom data and PET/CT ( ¹⁸F-fluoride and ¹¹C-choline) scans of patients with metastatic prostate cancer. Data are reconstructed with three different ordered subset (OS) methods: OSEM, OS Maximum A Posteriori One Step Late (OSMAPOSL) with the Median Root Prior and OS Separable Paraboloidal Surrogate (OSSPS) with the Quadratic Prior. These three algorithms are available with the open source Software for Tomographic Image Reconstruction (STIR, http://stir.sf.net). We investigate the SUVmax and SUVmean for each of the three reconstruction methods and how they change over sub-iteration. Our findings show that the subset limit cycle behavior does not affect the comparison of the SUV estimates for identical acquisition. On the other hand fine parameter tuning is necessary to achieve better tradeoff between convergence, bias and noise.