An optimized maximum likelihood approach to positron emission tomography image reconstruction

Abstract

In the maximum likelihood estimation (MLE) approach to reconstruction of emission tomography images, the iterative expectation-maximization algorithm is used. After too many iterations, artifacts appear. Here, optimized MLE reconstructions are obtained using a data splitting, cross-validation stopping rule approach. Reconstructions are postfiltered. Pixel-by-pixel statistics of reconstructions of real positron emission tomography data from fluorodexiglucose brain studies are tabulated for both the MLE and filtered backprojection approach. For regions where the average emission rate is low, MLE reconstructions had substantially lower standard deviations. For high activity regions, reconstructions from both methods had similar standard deviations. Hence, the MLE approach is expected to give better estimates of low activity regions. In the cross-validation approach, data can be split into more than two parts. The advantage of splitting into more than two parts is demonstrated for simulated data. >