List-Mode EM Algorithms for Limited Precision High-Resolution PET Image Reconstruction

摘要

Maximum likelihood (ML) list-mode expectation maximization (EM) reconstruction for positron emission tomography (PET) permits all acquired information to be used directly in the reconstruction process without preprocessing. This feature is particularly useful for high spatial and temporal resolution PET applications, such as high-resolution small-volume imaging, dynamic studies, and motion correction. However, the often substantial quantity (gigabytes) of list-mode data results in long reconstruction times and, unless appropriate measures are taken, bias due to limited machine precision. The use of subsets of list-mode data offers notable reduction in computing time (at least an order of magnitude), and this work shows that using subsets also overcomes the bias problem encountered in EM reconstruction on precision-limited computational platforms. Reconstruction performance with and without subsets for both ML and non-ML methods are compared in this article. Whereas simulated 2D data sets indicate increased variance in reconstructed voxel values through use of non-ML subset methods, measured 3D list-mode data show the highly accelerated non-ML subset methods produce results that are hard to visually differentiate from those of the ML algorithms (for the common case of regularization by stopping before reaching the ML estimate).