Improved attenuation correction in PET/MRI by combining MR image segmentation and joint estimation approaches

150 Objectives The goal is to improve MR-based PET attenuation correction in PET/MRI by combining two complementary approaches: 1) constructing an attenuation map based on MR image segmentation and 2) jointly estimating attenuation and activity maps from PET emission data. We hypothesize that the convergence of the two approaches enables us to meet challenges of addressing bone/lung/implant attenuation and segmentation errors. Methods A whole-body MR-based attenuation map (MRAM) is generated by segmenting Dixon fat-water images into 4 tissue classes (air, lung, fat and water) and then assigning predetermined attenuation coefficients to each class. In the MRAM, bone/implant components and patient-specific lung attenuation are not considered and classification/registration errors may exist. To address these problems, uncertain regions, likely to include bones/implants and lungs, are identified in MR images and then the attenuation coefficients in the identified regions are updated by estimating them jointly with an activity map from TOF PET data. In this study, regions likely to include lungs, ribs and/or vertebrae are identified in MR images by automatic procedures and then their attenuation values are updated by the joint estimation approach. Results For evaluation, a tri-modality (PET/CT and MR) clinical dataset was used. In an automatically identified vertebral region, the selective update by joint estimation improved MR-based attenuation coefficients when compared to reference CT-based attenuation coefficients (p=0.001, paired-sample t-test); a voxel-wise correlation coefficient of attenuation coefficients with the reference values increased from 0.58 to 0.84. A similar improvement was observed for an alternative region automatically identified as likely to include lungs, ribs and vertebrae (p Conclusions The proposed method of selectively updating MR-based attenuation maps by joint estimation demonstrated a promise to extract bone and patient-specific lung attenuation and address segmentation errors.