Combining different variance reduction approaches for PET image reconstruction

A variety of approaches have been proposed to reduce the variance in reconstructed PET images. In this work, we assess the effect of different combinations of variance reduction techniques on the quality of reconstructed images. These methods include MLEM with early termination, MLEM with post-smoothing, MAPEM and MLEM with inclusion of a convolution matrix prior to the system matrix. Different combinations of these methods have been implemented and tested on a raclopride 2D simulation. Evaluations were based on assessing single pixel values in the striatum region excluding the edges, since it is of interest for finding parametric images (e.g. binding potential). The results indicate that MAPEM regularization with the inclusion of a convolution matrix prior to the system matrix with or without post-smoothing is able to perform better than other regularization strategies used in isolation. Moreover, the analysis performed in this paper shows that the inclusion of a convolution matrix within MAPEM also reduces the sensitivity of the method to the regularization hyperparameter.