Computational analysis and improvement of SIRT

Area of Science:

• Medical Imaging
• Computational Science
• Image Reconstruction

Background:

• Iterative X-ray computed tomography (CT) algorithms offer high-quality imaging but face computational challenges, particularly at high resolutions.
• The Simultaneous Iterative Reconstruction Technique (SIRT) is a widely used iterative algorithm for image reconstruction.
• Optimizing iterative algorithms is crucial for practical applications, especially in resource-intensive fields like micro-CT.

Purpose of the Study:

• To develop an accelerated and more efficient version of the SIRT algorithm for X-ray CT.
• To introduce an automated method for determining the optimal relaxation parameter in SIRT.
• To improve the memory efficiency and data communication aspects of SIRT, particularly for distributed-memory systems.

Main Methods:

• An eigenvalue-based scheme was developed to automatically determine a near-optimal relaxation parameter for SIRT.
• A modified preconditioning approach was integrated into SIRT to solve weighted least squares problems, resulting in the PSIRT algorithm.
• The performance of PSIRT was evaluated using experimental residual norm and timing data from cone-beam micro-CT mouse data.

Main Results:

• The eigenvalue-based scheme accelerated SIRT convergence, requiring approximately half the number of iterations.
• The PSIRT algorithm demonstrated a smaller memory footprint and reduced data communication needs in distributed implementations.
• Experimental results validated the efficiency gains and performance improvements of the proposed PSIRT method.

Conclusions:

• The developed eigenvalue-based scheme effectively optimizes SIRT, significantly reducing computational demands for high-resolution CT.
• The PSIRT algorithm presents a more efficient and memory-conscious alternative for iterative image reconstruction in X-ray CT.
• This work contributes to making advanced CT imaging techniques more computationally feasible and accessible.