OBJECTIVE: We propose a statistical stopping criterion for iterative reconstruction in emission tomography based on a heuristic statistical description of the reconstruction process. METHODS: The method was assessed for MLEM reconstruction. Based on Monte-Carlo numerical simulations and using a perfectly modeled system matrix, our method was compared with classical iterative reconstruction followed by low-pass filtering in terms of Euclidian distance to the exact object, noise, and resolution. The stopping criterion was then evaluated with realistic PET data of a Hoffman brain phantom produced using the GATE platform for different count levels. RESULTS: The numerical experiments showed that compared with the classical method, our technique yielded significant improvement of the noise-resolution tradeoff for a wide range of counting statistics compatible with routine clinical settings. When working with realistic data, the stopping rule allowed a qualitatively and quantitatively efficient determination of the optimal image. CONCLUSIONS: Our method appears to give a reliable estimation of the optimal stopping point for iterative reconstruction. It should thus be of practical interest as it produces images with similar or better quality than classical post-filtered iterative reconstruction with a mastered computation time.
OBJECTIVE: We propose a statistical stopping criterion for iterative reconstruction in emission tomography based on a heuristic statistical description of the reconstruction process. METHODS: The method was assessed for MLEM reconstruction. Based on Monte-Carlo numerical simulations and using a perfectly modeled system matrix, our method was compared with classical iterative reconstruction followed by low-pass filtering in terms of Euclidian distance to the exact object, noise, and resolution. The stopping criterion was then evaluated with realistic PET data of a Hoffman brain phantom produced using the GATE platform for different count levels. RESULTS: The numerical experiments showed that compared with the classical method, our technique yielded significant improvement of the noise-resolution tradeoff for a wide range of counting statistics compatible with routine clinical settings. When working with realistic data, the stopping rule allowed a qualitatively and quantitatively efficient determination of the optimal image. CONCLUSIONS: Our method appears to give a reliable estimation of the optimal stopping point for iterative reconstruction. It should thus be of practical interest as it produces images with similar or better quality than classical post-filtered iterative reconstruction with a mastered computation time.