Abstract
581
Objectives This work assesses the performance of both a monotonic and a new non-monotonic algorithm for maximum likelihood (ML) PET image reconstruction. The recently proposed non-monotonic ML (NMML) algorithm has shown promising results in terms of convergence rate, compared with conventional MLEM. However, the more established preconditioned conjugate gradient (PCG) algorithm can also exhibit favourably fast convergence towards the ML solution, hence the need for this quantitative performance assessment of these contrasting approaches to reconstruction.
Methods The fast PCG algorithm and the recently proposed NMML method were implemented for the first time on the High Resolution Research Tomograph (HRRT). Their performance was evaluated using phantom data, 2D simulations as well as a [11C]DASB human dataset (Talbot PS et al. 2009)
Results Initial results on 2D simulated data show that both PCG and NMML achieve dramatically faster convergence compared to MLEM. The equivalent increase in the objective function after 90 iterations of NMML is achieved after 1017 iterations for MLEM, (i.e. x11 faster). Additionally, after 64 iterations both gradient methods demonstrated almost similar improvements in bias of up to 2% and 67% for high- and low-count regions respectively, compared to MLEM. However, a critical issue that needs to be addressed is the erratic behaviour of NMML, which can potentially affect the reconstructed image.
Conclusions The recently proposed NMML algorithm has been implemented for real and simulated HRRT data, and compared to the more established PCG method. Early results indicate similar convergence rates, but with slightly reduced errors in the NMML reconstruction. Current work concerns a full quantitative assessment of relative performance and stability for neuroreceptor studies with [11C]DASB