An Event-Driven Subsetting Approach for Accelerated PET List-Mode Reconstruction

PET iterative reconstruction typically employs subsetting to reduce computation time. List mode PET data provides flexibility for subsetting approaches, but optimization of list mode subsetting has not been fully explored. In this article, a new event-driven subsetting approach is described and tuned for brain imaging. Instead of the traditional “number of iterations” and “number of subsets” approach, the proposed method focuses on the “number of image updates” and “events per update.” The ability to change the number of events per update throughout the course of reconstruction provides acceleration opportunities, by first using small subsets to take rapid steps toward convergence and then increasing subset size to minimize subset-induced noise. We tested multiple parameterized approaches for variable-sized subsetting and found that a gradual increase of subset size provided stable acceleration. The number of projected events was reduced by nearly 50% as compared to constant subset size, while achieving equivalent image quality and mean squared error (MSE)-based convergence metrics. Tuning was performed for brain imaging on the GE SIGNA PET/MR, but the approach can be adapted for other imaging scenarios and scanners. The reconstruction acceleration achieved enables clinical translation of computationally complex list-mode reconstruction algorithms, such as event-based motion correction.