Development and Implementation of Fully Three-Dimensional Iterative Reconstruction Approaches in Spect with Parallel, Fan- And Cone-Beam Collimators.

As a noninvasive molecular imaging technique, Single Photon Emission Computed Tomography (SPECT) provides the distribution of radioactivity through reconstructing a projection dataset and allows us to visualize 3-dimentional (3D) functional information rather than anatomical information within a given organ. In this dissertation, three fully 3D iterative expectation maximization (EM) algorithms including maximum likelihood EM, ordered subsets EM and maximum a posteriori EM are studied in a variety of SPECT imaging systems with parallel-, fan-, and cone-beam collimators. To improve reconstructed image quality with quantitative accuracy, volumetric system models including a strip-area system model (SASM), a fan-volume system model (FVSM) and a cone-volume system model (CVSM) have been developed and implemented for parallel-, fan-, and cone-beam collimators, respectively. These volumetric system models accurately describe the photon detection process due to system geometric factors, and provide more precise detection probability than other simple approximations such as a line-length system model (LLSM). Besides the novel development in system models, we have introduced almost all degrading factors into SPECT and investigated one-ray, two-ray or four-ray attenuation correction, distance-, and angle-dependent detector sensitivity and resolution correction, and scatter correction. To suppress reconstruction noise in images, many regularization methods have also been compared and the 3D total variation method is the one most employed. Then, the performance of any new development has been evaluated both qualitatively and quantitatively through reconstruction of Monte Carlo simulations, experiments and clinical scans of numerical phantoms, physical phantoms, and patients. No matter what collimator is used in the projection acquisition, all results demonstrate that our new fully 3D iterative reconstruction approaches outperform the filtered backprojection (FBP) and EM algorithms with LLSM in terms of accuracy, bias, contrast-to-noise ratio, resolution, uniformity, and so on. Diagnostic performance of brain SPECT has also been evaluated on a group of patients and much higher sensitivity is observed in new approaches than that in FBP and EM-LLSM. To reduce the computational cost of our new reconstruction approaches, parallelized reconstruction using message passing interface has been proposed and studied on a cluster computer. Finally, different from widely used pixel-based reconstruction, a method for nonuniform image sampling iterative reconstruction in SPECT has been developed and investigated, in which a content adaptive singularity-based triangular mesh (2D) is employed to represent images. Preliminary results indicate that mesh-based reconstruction is an efficient and promising approach with respect to image quality, computational cost, and potential in motion tracking.