Influence of Time-of-Flight Kernel Accuracy in TOF-PET Reconstruction

• Main Authors: Daube-Witherspoon, M.E., Suleman Surti, Matej, S., Werner, M., Shridhar Jayanthi, Karp, J.S.
• Format: Conference Proceeding
• Language: English
• Subjects: Accuracy; Background noise; Computed tomography; Discrete event simulation; Image reconstruction; Image resolution; Imaging phantoms; Kernel; PET reconstruction; Positron emission tomography; Time-of-flight PET; Timing
• ISSN: 1082-3654; 2577-0829
• DOI: 10.1109/NSSMIC.2006.354230

In the reconstruction of data from time-of-flight (TOF) PET systems, the timing resolution is assumed to be known accurately; in iterative reconstruction, it is included in the system model, typically as a Gaussian function. The width of the reconstruction TOF kernel is taken to be equal to the measured coincidence timing resolution (tau). If tau changes (e.g., as a function of count rate), the TOF kernel used in reconstruction will not accurately model the measured data. The goal of this work was to assess the effect of using an inaccurate value of tau in the reconstruction TOF kernel. The Alderson phantom with 10-mm hot spheres was imaged on the Philips Gemini TF PET/CT system (585-ps intrinsic timing resolution). Distributions of hot spheres in 27- and 35-cm diameter warm cylinders were also simulated (trues only) for 300- and 600-ps timing resolutions. The data were reconstructed using up to 20 iterations of TOF-OSEM with 20 chronological subsets. For simulated data, reconstruction with a TOF kernel 10-25% narrower than tau led to a 3-12% decrease in contrast for all sphere sizes at the same background noise level because sphere events were misplaced due to the finite timing accuracy. Using a reconstruction TOF kernel 10-25% wider than tau resulted in a 3-7% increase in contrast. Even wider reconstruction kernels led to still higher contrasts for kernel widths up to 2-3 times tau, although at the cost of increased reconstruction time and slower convergence rates. For kernel widths greater than 3tau, the performance decreased, as the reconstruction approached the non-TOF algorithm. These trends were also observed with the measured data, although much less pronounced. The results indicate that contrast/noise performance is fairly insensitive to inaccuracies in the reconstruction TOF kernel, provided that it is not narrower than the actual timing resolution of the data.