Information-Adaptive Denoising for Iterative PET Reconstruction

Quantitative accuracy and thus diagnostic precision in Emission Tomography is impaired by the inherent random characteristics of the data acquisition leading to statistical image noise. Edge preserving spatial variation regularized iterative image reconstruction approaches require case-specific control parameter adaptation for optimized contrast-vs-noise tradeoff. For MLEM reconstruction, we propose and evaluate iRDF which automatically adapts RDP edge preservation parameters according to local image noise and PET data characteristics. In order to distinguish between clustered noise spots and small tumors, we introduce hot-spot artifact correction. The proposed method was evaluated using NEMA IQ phantom data as well as clinical patient data. After initial iRDF base parameter tuning, results showed that iRDF maintained similar image quality regardless of statistics without requiring manual parameter tuning, in contrast to e.g. RDP. With NEMA-IQ phantom data, local image variance was reduced to ~33%, while contrast of small spheres could be mostly preserved compared to nonregularized OSEM. Using a quarter of the originally acquired list-mode data, a noise decreased to ~22% while SUV-max has been reduced to ~75% of OSEM-based results. NEMA phantom and clinical data showed improved signal-recovery-to-noise ratios, leading to an overall ~3 times higher feature detectability especially in small lesions. Finally, the processed examples illustrate the effectiveness of the proposed hot-pixel artefact correction. The proposed auto-adaptive iRDF regularization demonstrates high potential to reduce the burden of prior parameter tuning. It realizes a trade-off between feature contrast and image noise on both local and global scale. According to increased noise robustness at different count statistics, iRDF can be considered an interesting alternative especially for low-dose PET imaging applications.