Aarhus University Seal / Aarhus Universitets segl

Lars Poulsen Tolbod

Bayesian Penalized Likelihood Image Reconstruction (Q.Clear) in oncology: Effects of contrast and count-rate dependence

Research output: Contribution to conferencePosterResearchpeer-review

Objectives: Point-spread function PET reconstruction (OSEM-PSF) has been shown to lead to non-physical signal recoveries for small lesions, due to Gibb’s ringing, if proper post-reconstruction filtering is not performed [1]. Recently, a Bayesian Penalized Likelihood Image Reconstruction algorithm with PSF (Q.Clear, GE Healthcare) has been introduced as an alternative to OSEM-PSF with better image convergence and activity-dependent noise regularizing. Q.Clear utilizes a noise penalty with both relative and absolute difference terms, which controls image noise locally with varying degrees of local smoothing. The strength of this noise penalty is controlled through the operator-set beta factor and the vendor recommended setting is beta=350-400 for oncological 18F-FDG PET. This is a different paradigm compared to conventional OSEM-PSF with post-reconstruction filtering, where image smoothing rather than noise is kept constant, and may have implications for the ability to suppress PSF artefacts and, thus, the quantitative accuracy of the PET images. To study the signal recovery features of Q.Clear as well as suppression of PSF artefacts at different contrast levels and total frame count rates, we examined scans of a clinically relevant phantom with small sub-centimeter lesions in direct contact with a background at varying lesion-to-background ratios. Methods: A wall-less phantom was prepared with homogenous spheres made of 18F-FDG dissolved in a gel with diameters ranging from 3 to 12 mm. The phantom consisted of two sets of spheres: one was suspended in 11C containing gel, providing a hot local background, whereas the other was suspended in a cold gel. The phantom was scanned for 210 min on a GE Discovery 690 PET/CT and reconstructed in 10 min frames resulting in sphere-to-background ratios (contrast) for the first set of spheres ranging from 0.8 to 35. The frame total count rate ranged more than one magnitude. Images were reconstructed either with Q.Clear (beta range 100 to 800) or OSEM-PSF. Time-of-flight was enabled in all reconstructions. The signal recovery of 4 common SUV metrics was evaluated (SUVmax, SUVmean, SUVpeak and SUV50bg). Results: Signal recoveries were monotonically dependent on sphere diameter regardless of contrast, reconstruction or count rate for SUVmean and SUVpeak. However, SUVmax and SUV50bg were sensitive to PSF artefacts and exhibited a non-monotonical dependence on sphere diameter with maximum recovery for the 8 mm sphere for Q.Clear with beta factors <=400 to 800 depending on contrast. Higher contrast required a higher setting of the beta factor. Similarly, SUVmax and SUV50bg exhibited decreasing signal recovery with decreasing frame count rate for constant contrast, due to the absolute count rate dependence of the penalty term. Conclusions: For images reconstructed with Q.Clear, signal recovery for small lesions was found to depend on both local contrast and total frame count rate. For 18F-FDG, for low contrast regions, a beta factor of 400 reduced the Gibb’s edge artefact, however, for regions with higher contrast a higher beta factor was need. The results imply that optimization of the Q.Clear penalty strength parameter depends on the organ imaged, the nature of the tracer as well as the injected dose regime. [1] Munk, O. L., Tolbod, L. P., Hansen, S. B., & Bogsrud, T. V. (2017). Point-spread function reconstructed PET images of sub-centimeter lesions are not quantitative. EJNMMI Physics, 4(1), [5].
Original languageEnglish
Publication year2019
Publication statusPublished - 2019
EventSociety of Nuclear Medicine and Molecular Imaging: Annual Meeting 2019 - Anaheim, United States
Duration: 22 Jun 201925 Jun 2019


ConferenceSociety of Nuclear Medicine and Molecular Imaging: Annual Meeting 2019
CountryUnited States

See relations at Aarhus University Citationformats

ID: 159034567