Abstract
1358
Introduction: Motion of the head during PET brain imaging studies results in burring and loss of spatial resolution. These movements are common and can be due to unintentional events, such as sneezing or coughing, or result from patients who have difficulty remaining still. Typically, head motion is corrected by splitting the acquisition into multiple frames, identifying the time that motion occurred, and throwing out data before or after the motion. Alternatively, one can register dynamic images from individual frames and average these images, but this suffers from suboptimal statistics and physics corrections. In this work, we propose a data-driven approach, using image-based motion estimation techniques on dynamic PET images and apply list-mode event-by-event reconstruction to compute a single data-driven motion-corrected (DDMoCo) image.
Methods: IRB-approved retrospective analysis was performed using clinical 18F-FDG PET/MR brain studies acquired on a Signa PET/MR scanner (GE Healthcare). Data used for this study included a 25-minute PET bed and vendor-provided MR-based attenuation correction scans. Dynamic PET images were generated without any corrections (e.g., scatter and attenuation) using a time-of-flight (TOF) list-mode reconstruction at varying time intervals (150s, 30s, and 5s) spanning the full 1500s PET scan. From the dynamic reconstructed images, image-based motion estimate was performed using the Analysis of Functional Images (AFNI) tool to compute rigid-body motion estimates. Motion parameters were then provided as inputs to an offline list-mode reconstruction tool [Spangler-Bickell, Khalighi, Hoo et al. (2018). IEEE TRPMS]. Images with and without MoCo were compared using peak signal-to-noise ratio (PSNR) and root mean squared error (RMSE).
Results: The use of DDMoCo provided visually improved PET images in all cases. In subjects with minimal motion, use of DDMoCo provided subtle improvements in image quality. For such subjects there was little difference between motion estimates at 150s, 30s, or 5s time intervals. In one subject with a single abrupt motion period, a 3-minute PET scan was compared to the quiescent period immediately before the motion event to a 3-minute PET scan during which motion occurred. With the application of DDMoCo vs no correction, besides having greatly improved image sharpness and visual quality, PSNR was 42.1 vs. 39.9 (5.5% higher) and RMSE was 438 Bq/cc vs. 557 Bq/cc (22% lower) demonstrating better quantitative image performance. Discussion/Conclusions: DDMoCo provided robust estimation of motion using image-based estimation from uncorrected dynamic PET reconstructions. For subjects with minimal motion, dynamic images with 30-150s temporal resolution appeared to adequately sample the observed motion, which may be typical for routine clinical studies. Motion estimation at a temporal resolution of up to 5s was demonstrated in a subject with a large, abrupt motion, showing improved visual image quality and more accurate quantification compared to non-moving data. A major advantage of this approach over image-based registration and averaging is that image SNR can be maximized since all counts are used in a single list mode reconstruction. One limitation of DDMoCo is the additional computation time necessary to reconstruct many short duration frames; however, this process could be parallelized and performed offline at the desired temporal resolution. Compared to other methods of motion estimation, DDMoCo requires no additional hardware or integration with the scanner system and could be readily implemented into a clinical environment as a tool to compensate for minimal to moderate head motion.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.