Evaluation of Penalized Maximum-Likelihood PET Image Reconstruction for ROI Quantification

Region of interest (ROI) quantification is one of the important tasks in positron emission tomography (PET) imaging. Statistical reconstruction methods based on the penalized maximum-likelihood (PML) principle have been developed to improve image quality over early stopping with EM-based methods and analytic methods. The penalty function that controls spatial smoothness plays an important role in PML reconstruction. The commonly used quadratic penalty often oversmooths sharp edges, while the edge-preserving penalty functions produce sharp intensity transitions across boundaries at the expense of potential image blockiness. Such blocky images limit the clinical widespread use of edge-prserving penalties. Recently patch-based edge-preserving penalties have been proposed for the purpose of preserving sharp edges and ensuring local smoothness while reducing image blockiness at the same time. The benefits of these penalty functions on clinical tasks had not been fully examined. In this work, we analyzed the ROI quantification performance for pairwise and patch-based penalties and also compared them with the standard ordered subsets expectation maximization (OSEM) reconstructions. We acquired a standard NEMA phantom scan and multiple whole-body patient scans with real lesions and also inserted lesions into the whole-body scans. The hot and cold sphere contrast recovery coefficient (CRC) versus the background variability was used to evaluate performance on the NEMA phantom. Mean lesion standardized uptake value (SUV) versus liver background variability was used for quantifying patient data. The results showed that the patch-based edge-preserving penalty achieved better quantification performance compared to other pairwise penalties and conventional OSEM.