Improving the image quality-quantification trade-off with AI-based PET denoising
Reducing injected activity or scan time only pays off if image quality and quantification hold. Using NEMA IEC phantom data from four scanner platforms, we show what to expect when NUCLARITY, Nuclivision's CE-marked deep learning denoiser for PET, is applied to half-count scans, which post-filter settings keep the result EARL 2-compliant, and how to find the count floor of your own scanner.
What is inside this white paper?
A practical, phantom-based guide for medical physicists and nuclear medicine department heads who want to reduce injected activity or scan time without losing quantification or image quality.
Why Gaussian post-filtering removes noise but also blurs small lesions, and how deep learning denoising shifts that trade-off in a different direction. Multi-centre NEMA IEC phantom data from four scanner platforms: lower background noise, higher contrast-to-noise ratio, and recovery coefficients that stay within EARL2 limits even at lower count acquisitions.
Why the count floor depends on scanner sensitivity, with a direct comparison between a high-sensitivity SiPM system and a lower-sensitivity PMT system. It shows where the limits are as well as the gains.
A practical, procedure to find the minimal post-filter that keeps your NUCLARITY-enhanced scans EARL2-compliant.
.png)
