RT Journal Article
SR Electronic
T1 Super-resolution 3-D PET reconstruction for simultaneous PET/MR
JF Journal of Nuclear Medicine
JO J Nucl Med
FD Society of Nuclear Medicine
SP 1367
OP 1367
VO 60
IS supplement 1
A1 Yanis Chemli
A1 Paul Han
A1 Chao Ma
A1 Georges El Fakhri
A1 Jinsong Ouyang
A1 Yoann Petibon
YR 2019
UL http://jnm.snmjournals.org/content/60/supplement_1/1367.abstract
AB 1367Objectives: PET images are characterized by their relatively low spatial resolution, caused by the physical phenomena occurring both at the emission and detection levels (e.g., positron range, inter-crystal scattering, etc.) as well as hardware limitations (e.g., finite dimensions of crystals). Super-resolution (SR) techniques can improve the spatial resolution by exploiting the spatial sampling information from multiple scans acquired with accurately known sub-resolution object shifts. The goal of this work was is to develop and evaluate an SR PET reconstruction algorithm for PET/MR, modeling all the physics of projection data formation and taking advantage of simultaneously acquired high-resolution MRI scans to precisely measure the object shifts. Methods: A Hoffman phantom filled with 2mCi 18F was scanned on a whole-body integrated PET/MR scanner (Biograph mMR, Siemens). A total of five 5-min PET list mode acquisitions were performed, each corresponding to a different phantom position and orientation. To measure the precise phantom location for each scan, a high-resolution MRI volume was simultaneously acquired using a Gradient-Recalled Echo sequence (TE=2.48ms, TR=6ms, flip angle=20°, in-plane resolution=1×1mm2, slice thickness = 1mm). A reference MR image was selected, and all other four MR images were rigidly co-registered to the reference to estimate the spatial shift parameters for each scan. The SR reconstruction method treats the sinogram data of a particular scan as a blurred, low-resolution, Poisson-distributed projection of an unknown high-resolution (HR) 3-D PET image shifted to a known position. The HR PET image that we seek to estimate depicts the object in the reference position and is estimated iteratively using sinogram data from all five scans. The SR reconstruction algorithm was implemented using OSEM, by modeling the estimated MR-derived shift parameters, image down-sampling, point spread function (PSF) blurring, forward-projection and detector sensitivity coefficients inside the PET system matrix. PET images were reconstructed with 3 different methods: (1) standard OSEM algorithm applied to the reference scan data (2x2x2mm3 voxel grid), (2) standard OSEM algorithm with PSF (OSEM-PSF) modeling applied to the reference scan (2x2x2mm3), (3) proposed SR algorithm applied to data from all five scans (1x1x1mm3). Each method used the same number of coincidence events to form the final images. No attenuation correction was performed. Iteration numbers were chosen to match noise levels across methods. Results: The proposed SR reconstruction method yielded PET images with visibly improved spatial resolution as compared to both OSEM and OSEM-PSF reconstruction, allowing for a better characterization of small cortical and subcortical brain phantom structures (Fig. 1). Line profiles confirmed the increase in spatial resolution for the SR image as well as improved correspondence with high-resolution MRI as compared to the conventional methods. Conclusions: This study indicates that super resolution PET reconstruction using simultaneously acquired MRI data for shift estimation is a promising way of improving PET image quality and resolution in PET/MR scanners. Acknowledgment: This work was supported in part by R01HL118261 and P41EB022544.