| CT image reconstruction methods can be grouped into two categories: analytic and iterative. The idea of iterative reconstruction (IR) is to compute projections based on an estimate to the original image, and iteratively compensate for errors between the computed and measured projection by updating the current image. IR is superior to analytic reconstruction in terms of image quality in many applications, such as in cases of noisy or incomplete projection data. The main drawback of the IR approach is its computational overhead. Parallel computing is a major means for acceleration of the IR process.; First, we applied the IR approach for dose reduction in peripheral quantitative CT (pQCT). We chose the expectation maximization (EM) method to reconstruct images from low-dose pQCT scans, and compared image quality to that obtained using the filtered backprojection (FBP) method. Low-dose projection data were numerically simulated and experimentally measured. Our results show that IR significantly outperformed FBP. With about 30% dose reduction, IR still generated acceptable image quality while FBP did produce distracting artifacts.; Then, we focused on IR acceleration by parallel computing. We parallelized EM, order-subset EM (OS-EM), simultaneous algebraic reconstructive technique (SART), and OS-SART on a Linux PC cluster. To achieve the same image quality obtained by the sequential algorithms in more than 45 minutes, the parallel OS-EM and OS-SART took less than 4 minutes in our representative simulation settings.; Furthermore, we studied the peer-to-peer (P2P) technology for IR, considering that decentralized parallel computing is potentially powerful and cost-effective. In the P2P scheme, the clients are directly connected to all other computing peers seamlessly and form a virtual computer. We proposed an Internet-based P2P system for IR, and tested its performance in numerical simulation.; Finally, we developed an IR-oriented GUI-based software package IterCT on a PC according to the software engineering principles. The implemented algorithms include EM, SART, OS-EM, OS-SART, and FBP for comparison. This first of its kind system handles parallel-beam, fan-beam and cone-beam data, has been posted as an open source shareware, and already well received by peers all over the world. |
