Effect of subsets on bias and variance in low-count iterative PET reconstruction

Statistical reconstruction produces bias in images reconstructed with low counts, such as short-duration frames in dynamic PET studies. While the bias is usually considered to be the consequence of the positivity constraint in MLEM reconstruction, there is no evidence of comparable bias observed with list-mode OSEM reconstruction (MOLAR). Simulation studies were performed of a human brain scan for the HRRT. Without scatter effects, there was only 0.5–4% negative bias in brain image reconstructed from as little as 0.2M events. With scatter included in the simulation and corrected in the reconstruction, additional bias was introduced, i.e., bias values were between −7% and −15% reconstructed from 0.2M prompt events (0.07M NEC). On the other hand, in both simulation and real phantom experiments, larger bias was found in low-count reconstructions if the events are divided into larger numbers of subsets. Compared to the case with only one subset, the bias and variability increased by 1–2% and 1–3% respectively when 30 subsets are employed in HRRT reconstruction from 0.2M simulated events. A similar tendency of increased bias was also observed when comparing 3 to 21 subsets in low-count reconstructions of a phantom scanned on the mCT. The negative bias in low-count OSEM reconstructions is presumably produced by too few counts being backprojected into the image at each subset, producing “holes”. This phenomenon is worsened by higher numbers of subset, i.e., fewer events in each subset. The results of this study indicated that image bias and noise in ultra low-count reconstructions with MOLAR can be potentially reduced by employing less subsets and more iterations.