%0 Journal Article
%A Ashrafinia, Saeed
%A Mohy-ud-Din, Hassan
%A Karakatsanis, Nikolaos
%A Kadrmas, Dan
%A Rahmim, Arman
%T Enhanced quantitative PET imaging utilizing adaptive partial resolution modeling
%D 2014
%J Journal of Nuclear Medicine
%P 371-371
%V 55
%N supplement 1
%X 371 Objectives Resolution Modeling (RM) in PET imaging models resolution degrading phenomena within reconstruction step, improving resolution and contrast recovery (CR). However, RM i) can generate Gibbs ringing artifacts and ii) may degrade reproducibility for small tumors. In this work, we introduce adaptive partial RM and study optimization in quantitative imaging tasks. Methods We considered lung tumor FDG PET imaging, with SUV images 60min post-injection simulated based on realistic clinically derived kinetic parameters. We modeled projection-space degradations (inter-crystal penetration & scattering, photon non-collinearity) including clinically realistic noise levels, and utilized the MLEM reconstruction algorithm. Our RM model includes a set of radial blurring kernels for different incident angles. Partial RM was obtained by introducing RM kernels with controllable parametrized (t) extents in-between no RM (t=0) and full RM (t=1) kernels. We assessed CR coefficients (CRC) curves, plus bias vs noise (spatial roughness, SUVmean coefficient of variation (COV)). Results No RM was outperformed by full RM with 40% reduction in bias at convergence. RM and partial RM had comparable bias vs. noise (whether defined as image roughness or SUVmean COV), as well as similar CRC curves. However, optimized (t=0.8) partial RM produced images with 15% reductions in SUVmean COV relative to full RM for matched SUVmean bias, and 12% reduction in SUVmean COV at matched SUVmax bias. The poorer reproducibility in full RM is due to increased inter-voxel correlations which lowers roughness (gives visual impressions of reduced noise) but actually degrades SUVmean reproducibility. Conclusions Partial RM achieves comparable CRC and image roughness relative to full RM, but shows significant reduction in SUVmean bias. We propose partial RM should be pursued as a very powerful and viable choice in quantitative task-based optimization including prognostication and treatment response assessment.
%U