Evaluation of Count-level Adaptive Image Domain PSF for PET Image Reconstruction

Abstract

Incorporating image-domain point spread function (PSF) modeling inside statistical reconstructions in emission tomography is capable of improving spatial resolution and tumor contrasts by modeling resolution degrading phenomena such as positron range, inter-crystal scatter and photon non-colinearity. However, such improvements do not necessarily translate into superior performance in clinical tasks and PSF modeling may introduce Gibbs-like ringing artifacts along edges due to the attempted recovery of high frequency information that is lost or heavily attenuated during data acquisition. In previous work we had proposed an approach for quantifying the visible portions of ringing artifacts and showed that for low-count datasets, overmodeling PSF kernels to a certain extent could help improve image quality while keeping visible ringing at a minimum. In this work, we develop a rule for automatically choosing the optimal PSF kernel width based on data quality and evaluate our method using a real patient study. We analyzed the performance of ROI quantification for the proposed PSF modeling approach and compared the results with standard, fixed-width PSF modeling using list-mode ordered subsets expectation maximization (OSEM) reconstructions. We acquired five whole-body patient scans with both real and inserted lesions. The contrast recovery coefficient (CRC) for inserted lesions or the mean lesion standard uptake value (SUV) for real lesions versus liver background variability was used for quantifying patient data. The results show that the proposed count-level-adaptive PSF modeling approach can achieve better quantification performance compared to standard PSF modeling without enhancing ringing artifacts.