Introduction

Simultaneous PET/MRI MoCo (e.g., the Biograph mMR BodyCOMPASS application) requires a companion dedicated MRI motion modeling scan, e.g., a golden-angle stack-of-stars sequence, throughout the entire PET acquisition duration for deriving 3D deformable MVFs ^1,2. This requirement reduces the overall PET/MRI efficiency. Therefore, a rapid motion modeling MRI scan is highly desirable. In this study, we tested the proposed PET MoCo approach by using MVFs obtained from a P2P-reconstructed short MRI scan (30-40 seconds) in a patient study. The deep learning P2P network was trained using only noisy images with artifacts to reconstruct 4D respiratory motion-resolved MRIs from highly under-sampled MRI data ³. We also examined whether the proposed method would provide comparable results to PET MoCo using MVFs obtained from a multi-coil nonuniform inverse fast Fourier transform (MCNUFFT)-reconstructed ⁴ long MRI scan (5-6 minutes) with various quantitative and qualitative PET metrics as endpoints.

Methods

The retrospective analysis was conducted using PET/MRI images acquired from six oncological liver patients for a total of 49 lesions using a 3T PET/MRI (Biograph mMR, Siemens Healthcare) scanner with IRB approval and HIPAA compliance. Imaging data were acquired using CAPTURE⁵ in sagittal orientation. The imaging parameters were TE/TR=1.66ms/3.5ms, FOV=360mm×360mm, slice partial Fourier=6/8, 50% slice resolution, # slices=96-120, voxel size =1.125×1.125×3 mm³. Two thousand radial spokes (Acq. time: 5:02min – 6:33min) and two hundred radial spokes (Acq. time: 30.2s – 39.3s) were acquired respectively. Eovist or Dotarem was administrated to the patients. Breath-hold VIBE MRI imaging was acquired and fused with PET. The k-space data were binned into 10 respiratory phases based on the 1D respiratory curve obtained from the CAPTURE. Two image reconstruction methods, including the MCNUFFT⁴ and the deep learning P2P, were used to obtain 4D respiratory motion-resolved MRIs. MCNUFFT was employed to reconstruct 4D MCNUFFT2000 MRI using all 2000 spokes. In contrast, the P2P reconstruction was applied to reconstruct 4D P2P200 MRI using only the first 200 spokes. In a previous study, the P2P network, adapted from the widely used U-Net architecture⁶, was trained using only noisy images with artifacts. Eight healthy participants were used for training and one for validation³. This previously trained P2P was then directly applied to the liver cancer patients’ data in the current study. A non-linear registration (ANTS)⁷ was employed on the MRIs to register all respiratory phases onto the end-expiration phase to obtain 3D deformable MVFs in the PET space. The difference in millimeters between two sets of in-plane MVFs [(u1, v1) vs (u2, v2)] was calculated based on the absolute error in flow endpoint (EE) used in Baker, Simon, et al⁸.
$$ EE = \sqrt{\left(u_{1} - u_{2}\right)^{2} + \left(v_{1} - v_{2}\right)^{2}} $$
Three sets of PET images were then reconstructed: 1) non-MoCo PET: 3D-OSEM reconstruction without motion correction; 2) MoCo₂₀₀₀: MCIR (motion-compensated image reconstruction)^9,10 using MVFs derived from MCNUFFT2000 MRI; and 3) MoCo_P2P200: MCIR using MVFs derived from P2P200 MRI. A non-linear registration (ANTS)⁷ was performed between the breath-held Dixon water image and the end-of-expiration CAPTURE MRI to reduce attenuation correction artifacts related to different respiratory phases. In the MCIR MoCo PET reconstruction, the magnitude of the MRI-derived respiratory motion curve was used to bin the simultaneously acquired PET list mode data into 10 sinograms based on the phase of respiratory motion. These 10 sinograms and their corresponding MRI-derived MVFs were then used in the iterative MCIR reconstruction to obtain MoCo PET images. Non-MoCo, MoCo₂₀₀₀, and MoCo_P2P200 were compared in terms of SUV_max, SUV_peak, SUV_mean, and lesion volume in the native PET space. Two radiologists blinded to the reconstruction methods scored lesions independently in terms of lesion conspicuity on a 5-point Likert scale.

Results

Images from phase 1 (end-of-expiration) of the 4D motion-resolved CAPTURE MRI acquisition, reconstructed using MCNUFFT2000 and P2P200, are shown in Figure 1 (upper row). The differences between MCNUFFT2000 MVFs and P2P200 MVFs are visualized in Figure 1 (lower row). The spaghetti plots of SUV_max, SUV_peak, SUV_mean, and lesion volume are shown in Figure 2. The analyzed lesions had significantly higher SUV_max, SUV_peak, SUV_mean, and significantly lower lesion volume on the MoCo₂₀₀₀ and MoCo_P2P200 than on the non-MoCo PET images. There was no statistical difference between MoCo₂₀₀₀ and MoCo_P2P200 with respect to SUV_max, SUV_peak, SUV_mean, or lesion volume. The mean rater conspicuity scores across reconstructions are summarized in a box plot shown in Figure 3. Figure 4 and 5 show non-MoCo, MoCo₂₀₀₀, and MoCo_P2P200 images from two liver cancer patients.

Discussion and Conclusion

Our results demonstrated that aligning breath-hold m map and free breathing PET data and incorporating MRI-derived MVFs in PET MoCo reconstruction improved SUVs and lesion conspicuity. Moreover, some liver lesions seen by both readers on the motion-corrected PET images were completely invisible to both readers on the non-motion-corrected PET images. Thus, motion-corrected PET images may enable radiologists to identify otherwise occult malignant liver lesions, potentially altering clinical management decisions.
Furthermore, we have shown that MoCo_P2P200 and MoCo₂₀₀₀ provided similar increases in various SUV metrics compared to non-MoCo, demonstrating that the MVFs derived using only a 30-40 seconds MRI motion model scan (i.e., MoCo_P2P200) was adequate. Our results showed that a short MRI motion modeling scan reconstructed using a deep learning P2P network can improve the overall imaging efficiency.

Acknowledgements

We acknowledge Siemens Healthineers for providing funding to this study.