Simulation-based evaluation of NEG-ML iterative reconstruction of low count PET data

In this study we propose the use of a new reconstruction scheme for low count data which consists in the reconstruction of the prompt events with the inclusion of the correction of the sources of contamination in the iteration process (OP-scheme) using the NEG-ML algorithm. The performances of this variant, hereafter called OP-NEG3D, is evaluated in low count data situations using simulated replicated data. Quantification results obtained with the filtered back projection algorithm (FBP3D), which is still the standard reconstruction method for quantitative dynamic PET brain studies, and conventional OP-OSEM3D are also shown for comparisons. The comparison tests include the VOI-based measurements of activity concentrations in large regions filled with high and low activity concentrations, the measurement of small activity differences in hot and cold regions using VOIs and using voxel-wise statistical analysis, and the activity recovery in small cold and hot inserts. The results indicated that OP-NEG3D allows an activity recovery in large cold regions as accurate as with FBP3D whereas OP-OSEM3D is biased due to the positivity constraint in the image space. OP-NEG3D shows better noise characteristics than FBP3D and allows a better detectability of differences between groups of scans and a better recovery of the activity in small regions than FBP3D. In conclusion, OP-NEG3D is a viable alternative to FBP3D for the reconstruction of PET data in the context of dynamic brain imaging.