Use of optimization transfer for enhanced direct 4D parametric imaging in myocardial perfusion PET

Objectives To investigate application of optimization transfer, via construction of a surrogate function, to achieve quantitative myocardial perfusion PET via direct 4D parametric imaging.

Methods Direct 4D parametric imaging in dynamic PET is commonly achieved via maximization of the Poisson log-likelihood (LL) function as expressed in the sinogram-domain. This optimization problem can be transferred to a surrogate function optimization task in the image-domain including iterative updates of the surrogate function, resulting in decoupling of the kinetic-parameter-to-image and image-to-sinogram relationships at every iteration. We applied this technique to direct 4D estimation of the forward transport K1 parameter (used to estimate myocardial blood flow). The algorithm incorporates preconditioned steepest ascent optimization of the Poisson log-likelihood surrogate function as constructed in every update. Clinical count-level simulations were performed for Rb-82 PET imaging based on real patient Rb-82 PET data on the GE Discovery RX PET/CT. Quantitative accuracy was assessed via noise (NSD) vs. bias (NMSE) trade-off curves for the entirety of the left ventricle (LV) as well as ten separate regions in the polar map.

Results The technique was seen to quantitatively outperform, not only conventional K1 parameter estimation performed post-reconstruction, but also direct 4D parametric estimation via sinogram-domain LL function optimization (over 15% reduction in noise; matched bias). Compared to the latter, a significant factor of 5 improvement in computational efficiency was also achieved due to the abovementioned decoupling between the sinogram and image space domains.

Conclusions Optimization transfer via construction of an image-domain surrogate function can significantly enhance quantitative and computational performance of direct parametric estimation as applied to myocardial perfusion PET imaging