Simultaneously Improving Accuracy and Precision within Dynamic Kernelized PET Reconstruction

Abstract

We propose two kernelized reconstruction methods for simultaneously improving accuracy and precision for dynamic PET imaging, followed by validations using 4D simulations. One of the proposed methods utilizes an effective kernel matrix which consists of an accuracy component that accounts for partial volume effect and a precision component which achieves 4D noise reduction (i.e. RBV-HYPR4D-K-OSEM), while the other method incorporates the standard resolution modeling within our previously proposed 4D de-noised reconstruction method (i.e. PSF-HYPR4D-K-OSEM). It was observed that the inclusion of the accuracy kernel improves the convergence rate in contrast recovery coefficient (CRC) for relatively small regions, whereas the inclusion of resolution modeling slows down the convergence rate. As compared to the standard OSEM with post filter, both proposed methods achieved better CRC vs noise trade off and mean absolute percent error across the time-activity curves, while RBV-HYPR4D outperforms PSF-HYPR4D in terms of accuracy.