A Robust No-Reference Quality Assessment Method for PET Imaging Based on Dynamic Thresholding, Poisson-Aware Noise Modeling, and Multi-Dimensional Block-Level Feature Fusion
Abstract
Purpose
No-reference positron emission tomography (PET) image quality assessment remains challenging due to Poisson-distributed noise, strong foreground–background contrast, and scan-parameter–dependent signal variations. This study aims to develop a PET-specific, adaptive NR-IQA method that is consistent with PET imaging physics and provides a robust, multidimensional image quality score applicable across variable scan conditions and clinical scenarios.
Methods
PET images are partitioned into fixed-size blocks and normalized prior to analysis. Activity, artifact, and noise detection are then performed on the normalized blocks. Blocks are classified as active when both cumulative activity and foreground ratio exceed predefined thresholds. A Poisson-aware noise model is introduced to identify noisy blocks, where block variance is used to quantify noise magnitude. Artifacts are detected using a Center–Surround Deviation measure. Scores from activity, noise, and artifact assessments are aggregated to produce a final image quality score, termed PET_PIQE. Method consistency was evaluated using Hoffman phantom images and clinical PET scans.
Results
For Hoffman phantom images with different quality levels, PET_PIQE showed strong correlation with effective image resolution (EIR) and coefficient of variance (CoV). PET_PIQE scores increased from 25.13 (EIR = 4.0, CoV = 14.23) to 32.66 (EIR = 4.7, CoV = 39.01) and 50.13 (EIR = 4.5, CoV = 79.26), reflecting progressive image degradation. In clinical PET scans acquired with durations of 10, 20, and 30 minutes at the same time point, PET_PIQE scores decreased with longer acquisition times (40.88, 39.97, and 37.36, respectively). Additionally, for scans performed at 1, 2, and 6 hours post-injection with identical scan durations, PET_PIQE scores increased over time (33.65, 40.34, and 45.56), consistent with declining image quality.
Conclusion
A PET-specific NR-IQA framework was developed that adapts to varying imaging conditions, incorporates Poisson noise characteristics, and evaluates multiple aspects of PET image quality.