Abstract
1839
Objectives The joint maximum likelihood reconstruction of activity and attenuation (MLAA) has recently gained attention, particularly in the context of emission-based attenuation correction in time-of-flight PET/MRI. Recently, we improved the performance of the MLAA algorithm using a multi-class Gaussian mixture model (GMM) constrained with MRI spatial and CT statistical information. In this work, we compare the performance of the proposed MLAA-GMM algorithm with standard 4-class MRI-based attenuation correction (MRAC) and reference CT-based AC.
Methods Five head and neck 18F-FDG patients were scanned on the Philips TF PET/MRI and Siemens mCT PET/CT scanners. Dixon fat and water MR images were coregistered to CT images. Standard 4-class MRAC maps were then derived by segmenting the MR images into background air, lung, fat and soft tissue followed by assignment of predefined attenuation coefficients. For MLAA-GMM, MR images were segmented into the same 4 classes plus an unknown MR low-intensity class encompassing cortical bones, air cavities and metal artifacts. To eliminate the mis-classification of bones into surrounding tissues, a co-registered bone probability map is included in the unknown tissue class. Finally, the GMM prior was constrained over known tissue classes of attenuation maps using uni-modal Gaussians parameterized over a patient population.
Results The results show that the MLAA-GMM algorithm outperforms the MRAC method by differentiating bones from air gaps and estimating more accurately the patient-specific attenuation coefficients of soft tissue and lungs. The MRAC method resulted in average SUV errors of -5.4 ± 12.0%, -7.4 ± 6.6% and -18.4 ± 7.9% in lungs, soft tissues/lesions and bones, whereas the MLAA-GMM algorithm reduced the errors to -3.5 ± 6.6%, -5.0 ± 5.5% and -10.2 ± 6.5%, respectively.
Conclusions The proposed algorithm MLAA-GMM outperforms the standard MRAC technique and is promising for deriving accurate attenuation maps in TOF PET/MR imaging.
Research Support This work was supported in part by the Swiss National Science Foundation under Grant SNSF 31003A-149957 and by the Indo-Swiss Joint Research Programme ISJRP 138866.