Evaluation of the Accuracy and Robustness of a Motion Correction Algorithm for PET Using a Novel Phantom Approach

We introduce the use of a novel physical phantom to quantify the performance of a motion-correction algorithm. The goal of the study was to assess a PET-PET image registration, the final output of which is a motion-corrected high-count PET image volume, a procedure called Reconstruct, Register and Average (RRA). Methods: A phantom was constructed using 5 ~ 2 mL Ge-68 filled spheres suspended in a water-filled tank via lightweight fishing line and driven by a periodic motion. Comparison of maximum and mean activity concentration and sphere volume was performed. Ground truth data were measured using “no motion.” With motion, five replicate datasets of 3-minute phase-gated data for each of three different periods of motion were acquired. Gated PET images were registered using a multi-resolution level-sets-based non-rigid registration (NRR). The NRR images were then averaged to form a motion-corrected, high-count image volume. Spheres from all images were segmented and compared across the imaging conditions. Results: The average center-of-mass range of motion was 7.35, 5.83, and 2.66 mm for the spheres over the three periods of 8, 6 and 4 seconds. The center-of-mass for all spheres in all conditions was corrected to within 1 mm on average using NRR as compared to the gated data. For the RRA data, the sphere maximum activity concentration (MAC) was on average 40.2% higher (-4.0% to 116.7%) and sphere volume was on average 12.0% smaller (-8.2% to 28.1%) as compared to the ungated data with motion. The RRA results for MAC were on average 70% more accurate and for sphere volume 80% more accurate as compared to the ungated data. Conclusions: The results show that the novel phantom setup and analysis methods are a promising evaluation technique for the assessment of motion correction algorithms. Benefits include the ability to compare against ground truth data without motion but with control of the statistical data quality and background variability. Use of a non-moving object adjacent to spheres in motion, the spatial extent of the motion correction algorithm was confirmed to be local to the induced motion and to not affect the stationary object.