Image properties of various ML-based reconstructions of very noisy HRRT data

The use of iterative image reconstruction algorithms with resolution modeling allows for reduced partial volume effect without noise increase. However, it is now recognized that EM-ML type algorithms are biased in very low counts images, in particular for cold regions. Alternative ML algorithms that allow for negative image voxels have been proposed to reduce the bias: NEG-ML of Nuyts et al and AB-ML of Byrne. The aim of this study is to evaluate the NEG-ML and AB-ML algorithms with respect to the EM-ML and 3DRP algorithms for human brain dynamic studies on the high resolution HRRT scanner. As the ground truth is not known, the idea is to distribute the list-mode data into several statistically independent gates of 2ms duration and to compare the summation of the reconstructions of each gate to the reconstruction of the entire list-mode. As both images are produced using the same data, differences are solely due to low count statistics effects. The average number of events per repUcate is selected by choosing the number of gates. A one hour FDG brain PET study was split into different numbers of gates (ranging from 2 to 360). The count statistics of each replica ranged from a 30min to a 10s acquisition, resulting in weakly to extremely noisy data. Two threshold values in the denominators of the NEG-ML equation were used: 1 as originally recommended and 10^-4 as frequently used for EM-ML. Arbitrary high bound values were used for the AB-ML algorithm. The mean activity concentration was measured in the gray and white matters, based on regions of interest drawn on a MRI of the same subject. As expected, no differences were found for 3DRP between the one hour acquisition and the sum of shorter acquisitions. EM-ML and NEG-ML with the lower threshold lead to an absolute bias ranging from 0.1 to 4.5% for 30min to 10s acquisitions. The bias was reduced to less than 0.5% for AB-ML and NEG-ML with a threshold of 1, for all acquisition durations. Convergence and noi- e properties were also studied. The AB-ML and NEG-ML algorithms are almost bias-free alternatives to EM-ML PET reconstruction of data with any noise level. The choice of the denominator threshold of the NEG-ML algorithm seems to play an important role.