Acceleration of list-mode expectation maximisation-maximum likelihood

Abstract

List-mode data preserves all sampling information from 3D PET imaging, and reduces storage requirements for multiple time frame acquisitions. List-mode EM-ML, which has been implemented in a number of forms (such as the EM algorithm for list-mode maximum likelihood, the FAIR algorithm and COSEM), is an obvious choice to reconstruct from such data sets when the statistics are low. However, these methods can be slow for large quantities of mode data, and it is desirable to accelerate them. This work investigates the use of subsets in combination with a relaxation parameter for 3D list-mode EM-ML reconstructions. Results show just two iterations through the list-mode data are sufficient to aid good quality reconstructions. Furthermore, if counting statistics are good, just one iteration may prove sufficient, opening the way for real-time iterative reconstruction.