A 4-D Iterative HYPR Denoising Operator Improves PET Image Quality

Abstract

There is an increasing demand for high spatial and/or temporal resolution dynamic PET images in research and clinical settings. Such images often have a low number of acquired counts per voxel, leading to poor signal-to-noise ratio, thus hampering quantitative accuracy and precision of image features. This can be obviated by a bias-free postprocessing denoising algorithm to improve precision while preserving feature accuracy. Highly constrained backprojection (HYPR) is a denoising algorithm that offers substantial denoising while preserving resolution using a 3-D composite image—usually a weighted sum over all dynamic frames. However, HYPR still introduces bias in frames where the feature contrast differs from that in the composite, and HYPR denoising is limited by the composite noise level. In this work, we extend the HYPR operator to be iterative and 4-D to minimize the potential mismatching contrast between the composite and frames. The initial composite is generated using regional averages instead of temporal sums to improve the level of denoising. Through phantom, simulation, and human studies, we demonstrate that this iterative 4-D HYPR (IHYPR4D) operator yields improved accuracy and precision compared to the traditional 3-D HYPR.