The new generation of PET scanners is approaching the resolution limit of the modality through the use of new scintillators, smaller detectors and 3D data acquisition. However the filtered backprojection (FBP) image reconstruction schemes used in most clinical settings are unable to fully realize the potential of these systems as they do not take account of the photon-limited nature of the data or the fact that the coincidence data do not represent perfect collimation between individual detector pairs. The thesis of this project is that PET image reconstruction is best performed within a statistical framework in which the physical and statistical properties of the observed data are accurately modeled. In the current funding period fast 2D and 3D Bayesian reconstruction methods were developed which combine accurate statistical and physical models of coincidence detection with image models designed to reflect the piece-wise smooth nature of in-vivo tracer distributions. These methods have been compared to standard clinical protocols and shown to produce improvements in resolution at matched background noise levels, quantitative accuracy, contrast recovery, and lesion detection using observers. During the proposed project period they will continue to extend the Bayesian reconstruction methods to a range of state of the art scanners: (I) whole body CTI EXACT HR+, (ii) the new CTI high resolution brain scanner with LSO based depth-of-interaction detectors, (iii) the simultaneous x-ray CT/PET scanner being developed jointly by CTI and the University of Pittsburgh, (iv) the UCLA microPET small animal scanner, and (v) the dedicated LSO-based PET mammography system under development at UCLA. They will use the factored matrix method to model positron range and photon-pair angular separation, geometric and intrinsic detector efficiencies, crystal penetration and scatter, and dead-time. Within the conjugate gradient Bayesian image reconstruction framework they will include a range of Gibbs priors (e.g. convex pair-wise smoothing, anatomically-based priors, spatially variant weightings for uniform resolution) and likelihood functions (e.g. Poisson, Gaussian approximations, modified Poisson models to account for the effects of randoms subtraction). They will include Ordered Subsets EM (OSEM) implementation options in the code for comparative purposes and to allow fast reconstruction. They will also implement a list mode likelihood variation of our methods. Optimized multithreaded implementations of these image reconstruction methods will be developed for use on single and multiprocessor UNIX workstations and Pentium-Pro servers. Performance of Bayesian, OSEM and optimized FBP reconstruction methods will be evaluated through studies of resolution, signal to noise measures, quantitation and lesion detection. Quantitation studies over small and irregularly shaped regions of interest that include consideration of partial volume effects will be performed using phantoms and combined autoradiography, MR and PET studies of small animals. Lesion detection will be studied using data from a realistic breast & thorax phantom and simulated lesions added to human data. Our optimized software will be made available to our collaborators and other interested research sites.
