PET image reconstruction using multi-parametric anato-functional priors

Abstract

The recent advent of clinical simultaneous PET-MR scanners has triggered an emerging paradigm toward multi-parametric data acquisition and image analysis with the aim of improving the diagnostic confidence of PET and MRI findings. One of the promising aspects of these intrinsically co-registered dual-modality systems is MR –guided PET image reconstruction given that the PET has a lower resolution and the reconstructed images usually suffer from noise or, in the case of resolution modelling, from Gibbs ringing artifacts. In this study, we investigated the utilization of multi-contrast MRI as well as PET information to guide PET image reconstruction, an unexplored aspect of multi-parametric PET-MR imaging that aims to address the pitfalls of conventional MR-guided PET image reconstruction methods owing to the complementary nature of the available data in these scanners. The current state-of-the art anatomical priors were investigated and presented in a unified framework and the capable priors were extended to the multi-parametric anato-functional priors. We studied the conventional local Tikhonov and total variation (TV) priors, and the anatomical priors such as Kaipio, non-local Tikhonov with Gaussian and Bowsher similarity kernels and a modified Burg joint entropy using extensive 3D realistic simulations and two clinical datasets of brain including [18F]florbetaben and [18F]FDG which have distinct tracer uptakes in the white and grey matter, respectively. The Gaussian kernels were calculated using both voxel- and patch-based feature vectors. Our simulation results showed that the Burg joint entropy far outperforms the conventional anatomical priors in preservation of PET unique lesions and reconstruction of functional boundaries which have corresponding MR anatomical boundaries. In addition, it was found that the multi-parametric extension of the priors leads to enhanced preservation of edge and unique features in PET images, and also an improved bias-variance performance. Consistent with our simulation results, the clinical image reconstructions showed that the Gaussian similarity kernels with voxel-based feature vectors, the Bowsher method and the Burg prior are the best performing priors among the studied anatomical priors and their multi-parametric extensions led to the improved recovery of the PET unique features. In conclusion, multi-parametric anato-functional priors provide a reliable solution to address the pitfalls of the conventional anatomical priors and are therefore likely to increase