PSF Overestimation Improves PET Image SUV Quantification

Objectives Point spread function (PSF) modeled PET image reconstruction (also known as resolution modeling) improves resolution and contrast, yet may degrade reproducibility and depict edge artifacts (reminiscent of Gibbs ringing). This study investigates the effect of incorporating varying PSF kernels, enabling under- and overestimation of the true PSF, for the potential of enhanced quantitative SUV estimation.

Methods In the context of oncologic FDG PET, we simulated 60min post-injection SUV images of XCAT anthropomorphic phantom (3min/bed) separately including six liver tumors of varying diameters (10, 13, 17, 22, 28 and 37mm). This was based on clinically estimated kinetic parameters from which time-activity curves were generated, followed by temporal integration. Normalization, attenuation, decay correction, and resolution degrading phenomena (including photon non-collinearity, inter-crystal penetration and scattering), as well as clinically realistic noise levels (100 realizations) were incorporated. We introduced controllable scaling factors to generate 20 analytically modeled blurring kernels (incorporating abovementioned physical phenomena with additional scaling in extent), including no/full PSF modeling, as well as 8 underestimated and 10 overestimated kernels. For performance comparison, we performed threshold-based tumor segmentation, and subsequently assessed contrast recovery coefficient (CRC), SUVmean bias, noise (spatial roughness, SUVmean and SUVmax coefficient-of-variability (CoV)), and averaged difference of max and min uptakes in every tumor.

Results For overestimated kernels, CRC was significantly improved, and noise (SUVmean CoV, SUVmax CoV and spatial roughness) vs. bias curves were also typically improved. The latter is because overestimated PSF kernels yield increased inter-voxel correlations, resulting in visually smoothed images and significantly lower spatial roughness (25% less for PSF vs. No PSF, and 43% less for overestimated PSF vs. no PSF, for clinically realistic number of iterations in each method used throughout this work). Furthermore, PSF modeling edge overshoots actually reduced SUVmean bias for the three smallest tumor (10% for true PSF and 15-20% for mid overestimated PSF), and resulted in comparable SUVmean performance in larger tumors. SUVmean CoV (as measure of reproducibility) for smallest tumors had a limited (5-10%) increase compared to true PSF, which was offset by more significant reductions in bias when overestimating. Averaged max-min difference as well as spatial roughness, used as substitutes for visual uniformity assessment (impacted by both presence of noise and edge artifacts), were actually reduced by 35% and 50% for true PSF and overestimated PSF in all six regions, due to overriding reduction in noise relative to enhancements in edge overshoots.

Conclusions Overestimated PSF yields significantly higher contrast for a range of tumors. For reasonable (not excessive) number of iterations, edge enhancement counterintuitively lowers SUVmean bias in small tumors, while inter-voxel correlations suppressing image roughness, and enhance uniformity, in all tumors, only slightly degrading SUVmean reproducibility in the smallest tumors. These results provide perspective on performance of vendor PET images, as some measure the PSF utilizing Ge-68 point sources (that have higher positron range than F18 as commonly used in the clinic), thus implicitly implementing overestimated PSF kernels. This, we suggest, may be enabling enhanced SUV quantitation, while we are exploring impact on measures of heterogeneity. Overall, we propose that overestimated PSF kernels can be pursued as powerful and viable choices in quantitative task-based optimization including prognostication and treatment response assessment.

• Download figure
• Open in new tab
• Download powerpoint