MR constrained simultaneous reconstruction of activity and attenuation maps in brain TOF-PET/MR imaging

The maximum likelihood estimation of attenuation and activity (MLAA) algorithm has been proposed to jointly estimate activity and attenuation from emission data only. Salomon et al employed the MLAA to estimate activity and attenuation from time-of-flight PET data with spatial MR prior information on attenuation. Recently, we proposed a novel algorithm to impose both spatial and statistical constraints on attenuation estimation within the MLAA algorithm using Dixon MR images and a constrained Gaussian mixture model (GMM). In this study, we compare the proposed algorithm with MLAA and MLAA_Salomon in brain TOF-PET/MR imaging.

A clinical FDG head PET/CT/MR dataset was used to simulate a 40M-count PET data acquisition with TOF resolution of 580 ps. In MLAA_GMM, Dixon MR images are segmented into outside air, fat/soft tissue classes and an MR low-intensity class corresponding to air cavities, bone and susceptibility artifacts. A mixture of 3 Gaussians (air, fat/soft tissue and bone) was used for the low-intensity class, while uni-modal Gaussians were used for other classes. Bias performance of the algorithms was evaluated against CT-based and 4-class MR-based attenuation correction methods.

Region-of-interest analysis of our simulations showed that the 4-class and MLAA algorithms result in –4.9% and –5.8% bias in soft tissue and –18.5% and –12.4% bias in bone, respectively. Inclusion of MR constrains in MLAA_Salomon and MLAA_GMM resulted in –6.6% and –4.1% bias in soft tissue and –16.1% and –13.0% in bone, respectively. It was found that the performance of MLAA_Salomon depends highly on the robustness of MR segmentation, particularly at air/bone interfaces.

The proposed approach effectively exploits MR prior information and produces attenuation maps that are spatially and statistically more consistent with true attenuation maps.