Analysis of MLEM and OSEM Reconstruction Algorithms for Different Geometries of PET Scanners

At the experimental heavy ion tumor therapy facility at the GSI Darmstadt in-beam PET is operated. The in-beam PET data should be reconstructed as fast as possible (ideally, in real time), because they allow monitoring irregularities in beam delivery. To find a fast reconstruction algorithm with acceptable quality, MLEM [1] and OSEM [2] algorithms were compared for full-ring and dual-head geometries of the PET scanner with different parameters of reconstruction. An activity distribution to be reconstructed was generated by the simulation routine. Quality assurance of reconstructions was performed by numerical and visual comparison of the reconstructed and generated activity distributions. Introduction The simulated activity was homogeneously distributed within a hollow sphere. Then it was reconstructed using the projection data by means of MLEM and OSEM algorithms. Results of these reconstructions were compared using the Mean Absolute Error (MAE) [3]. For the OSEM algorithm projection data were grouped into different number of subsets. Results for the full-ring PET scanner Reconstructions with OSEM 6, 8, 10 and 12 subsets were compared with reconstruction result of 50-th iteration of MLEM (Fig. 1). Best result in terms of MAE showed the reconstruction with OSEM 8 subsets. It achieved the quality of 50-th iteration of MLEM in 9 iterations. Figure 1: Mean Absolute Error for reconstructions within the full-ring PET scanner. Results for the dual-head PET scanner In case of a dual-head PET scanner MAE has a minimum at the 35-th iteration for MLEM and then increases. Finally MAE of MLEM and OSEM algorithms converges to the same value (Fig. 2). Figure 2: Convergence of MLEM and OSEM for the dual-head PET scanner geometry. The best reconstruction result showed the OSEM algorithm with decreasing number of subsets from iteration to iteration. It achieved the quality of the 35-th iteration of MLEM in 2 iterations (Fig. 3). Figure 3: Mean Absolute Error for reconstructionswithin the dual-head PET. References[1] L.A. Shepp and Y. Vardi, “ Maximum likelihood re-construction for emission tomography” , IEEE Trans.Med. Imag., vol MI-2, pp.113-122, 1982.[2] H.M. Hudson and R.S. Larkin, “ Accelerated ImageReconstruction Using Ordered Subsets of ProjectionData” , IEEE Trans., vol 13 No 4, pp.601-609, 1994.[3] J. Sheng and D. Liu, “ An Improved Maximum Like-lihood Approach to Image Reconstruction Using Or-dered Subsets and Data Subdivisions” , IEEE Trans.Nucl. Science Vol 51 No 1, pp. 130-135, 2004. ________________________________________*Work supported by GSI Darmstadt and EUalso with Technische Universitä t DresdenRAD-BIOPHYSICS-30