PET motion correction using real-time subspace-based MR imaging

Introduction: Image quality of PET reconstructions is degraded when motion occurs during the acquisition. Sources of motion include cardiac, respiratory and bulk motion. PET/MR offers a unique opportunity to estimate the motion using MR and correct for it during PET reconstruction. The conventional MR-based motion correction methods first group the k-space MR data into different motion phases (or bins) using navigators or physio signals and then reconstruct images at each motion phase for motion field estimation. However, the binning process assumes periodic motions and could fail when bulk or irregular respiratory/cardiac motion occurs. In this work, we propose to reconstruct dynamic MR images at high frame rate (100 ms/frame) from sparsely sampled (k,t)-space data for motion field measurement in real time. Our method is enabled by a subspace-based MR data acquisition and image reconstruction method.

Methods: To evaluate the effect of the proposed MR-based motion correction on PET image quality, an 18F-FDG-PET/MR scan was performed on a healthy subject using a PET/MR scanner (Siemens Biograph mMR). The acquisition time was 5 minutes starting 30 minutes after injection. To assess the effect of both respiratory and bulk motion, the subject was instructed to move twice during the scan. MR acquisition was performed using a spoiled gradient-recalled echo (GRE) sequence (TR/TE = 3/1.6 ms and 7 degree flip angle) with stack-of-stars radial sampling scheme. MR data was acquired with frame rate of 100 ms per frame, consisting of 3 lines for training and the rest for imaging. The spatial resolution was 1.9 mm in-plane and 5 mm through-plane (32 slices). Acquisition with the vendor-provided standard two-point Dixon sequence was performed for attenuation correction. MRI reconstruction from highly undersampled k-space data was performed using a subspace-based image model that leverages a unique property of dynamic MR images, known as partial separability [1]. Iterative reconstruction used ADMM [2] to minimize a cost function composed of a data fidelity term along with a low-rank and spatio-temporal TV penalty. Binning of the MR images was then performed retrospectively from the reconstructed dynamic MR images by first delimiting the bulk motion frames and tracking the position of the tip of the liver for each phase of the bulk motion. A total of 18 bins (6 bins for each of the 3 patient positions) was used. Motion estimation was performed on the binned MR images using B-spline deformation. Finally, PET reconstruction was performed on binned sinograms using the OSEM algorithm [3]. The imaging model for the PET acquisition included motion via a warping operator from the reference bin to each motion bin. 5 iterations and 12 subsets were used for the OSEM algorithm. No scatter correction was performed.

Results: Reconstructed MR images for different phases of the bulk and respiratory motion are shown on Fig.A along with the estimated liver position. Coronal and axial slices through the reconstructed PET volume are shown on Fig.B. The figure shows reconstructions without motion correction (NMC), using gating with 15% of the PET data (Gated) and with motion correction estimated from MRI (MC). The proposed method reduces motion blur artifacts as seen visually from Figs.B and C. Finally, the contrast to noise ratio was computed between the kidney and the surrounding background. The proposed MC method improves the CNR by a factor 2 over NMC and 4 over the Gated reconstruction.

Conclusions: The proposed MR acquisition and reconstruction method recovers high quality MR images in real-time (100 ms per frame) able to capture all types of patient motion. When used for motion correction in PET reconstruction, results demonstrate suppression of motion artifacts. Research support: T32EB013180, R01CA165221, R01HL118261, R21MH121812, R01HL137230 and P41EB022544

• Download figure
• Open in new tab
• Download powerpoint