Effect of ordered-subsets reconstruction on PET lesion detection

Objectives Iterative reconstruction in PET is computationally demanding, with block-iterative approaches such as ordered-subsets expectation-maximization (OSEM) commonly used to speed the reconstruction. These methods affect noise propagation and convergence, potentially degrading clinical performance. In this work, the effect of block-iterative reconstruction upon PET lesion-detection performance was assessed using a numerical observer and localization receiver operating characteristics (LROC) analysis.

Methods Twelve fully-3D anthropomorphic phantom experiments from the Utah PET lesion-detection database were retrospectively used, modeling whole-body FDG cancer imaging with distributed lesions (diam. 6-16mm) appropriate for LROC analysis. The data were reconstructed using maximum-likelihood (MLEM), OSEM, and rescaled block-iterative (RBI-EM) algorithms. All evenly-divisible subset sizes in azimuthal angle were considered (n=2, 4, 6, 7, 8, 12, 14, 21, 24, 28, 42, 84, 168). The channelized non-prewhitened observer was applied to all iterations of each case, and the Gaussian post-filter that maximized performance was identified. Tumor localization performance and area under the LROC curve (A_LROC) were then analyzed vs. number-of-subsets for both OSEM and RBI-EM, using MLEM as a basis for comparison.

Results Moderate subsetting (2-12 subsets) provided performance similar to MLEM (A_LROC = 0.39), whereas more aggressive subsetting significantly degraded lesion-detection performance (A_LROC = 0.18 – 0.37). Such degradation was less pronounced for RBI-EM than for OSEM.

Conclusions Within the limitations of this study, OSEM with 2-12 subsets accelerated reconstruction without adversely affecting lesion-detection performance. RBI-EM experienced less degradation than OSEM, potentially permitting somewhat more acceleration without loss of detection performance.

Research Support R01 CA107353