SU-E-J-14: A Novel, Fast, Variable Step Size Gradient Method for Solving Simultaneous Algebraic Reconstruction Technique (SART)-Type Reconstructions: An Example Application to CBCT

Abstract

Purpose: To develop a fast‐converging SART‐type algorithm and show clinical feasibility in CBCT reconstructions by combining the algorithmic enhancements with a parallel computing hardware (GPU). Methods: SART reconstructs a volumetric image by iteratively conducting volume projection and correction backprojection. However, this reconstruction problem can also be cast as a least squares problem for minimizing the volume projection errors with respect to the scanner projection data. This way, SART can be viewed as equivalent to a gradient method for minimizing the quadratic objective function f(x), with a fixed step‐size. Novelty in this work is that we implemented a simple yet much faster algorithm by computing a unique step‐size at each iteration. We applied this variable step‐size (VS)‐SART algorithm to numerical and physical phantoms for reconstruction. Furthermore, we accelerated the reconstruction by implementing the algorithm on NVIDIA GTX 295 GPU card. CBCT projections of CatPhan phantom were acquired from the Varian TrueBeam system. Results: We first compared SART and VS‐SART using Shepp‐Logan numerical phantom with 180 parallel‐beam projections. As the iterations progress, f(x) is asymptotically minimized for both algorithms but VS‐SART is found to converge much faster. Therefore, for a fixed number of iterations, VS‐SART commands superior image quality. In addition, compared with the FDK algorithm with 364 projections, our VS‐SART algorithm produces visibly equivalent quality CBCT image for CatPhan phantom with only 120 projections, in 12 iterations completed in 33 seconds. This is a factor of three dose reduction while maintaining the reconstruction time acceptable. Conclusions: By approaching SART reconstruction problem from a gradient method perspective, we enhanced the reconstruction speed significantly (i.e., less number of iterations). In addition, with GPU acceleration, the overall reconstruction time is achieved within a clinically viable range. We anticipate that the VS‐SART algorithm can be applied to CT, PET, and SPECT also.

Smoothed L0-Constraint Dictionary Learning for Low-Dose X-Ray CT Reconstruction

“…In addition, Fig.8 shows the NMAD curves of the human head slice phantom with different noise levels. It can be found that results produced by 1×10 5 photons and 2×10 5 photons are poor, 4×10 5 photons is better, and 6×10 5 photons is enough to produce very good results, then adding more photons may not show obvious improvement of reconstruction quality.…”

“…The analytic-based algorithms are derived from a continuous imaging model so that it really needs dense sampled projections following the Nyquist/Shannon sampling theorem [4] . For the reconstruction problem of incomplete projection data, algebraic algorithms like the simultaneous algebraic reconstruction technique (SART) [5] performs better compared to analytical algorithms.…”

Low-dose CT reconstruction via L1 dictionary learning regularization using iteratively reweighted least-squares

“…To solve the under-sampled reconstruction problem, algebraic algorithms transform the problem to a series of linear equations and the reconstructed image is acquired by the iterative method. SART (simultaneous algebraic reconstruction technique) [ 5 ] is one of the typical algebraic iterative methods. But the images reconstructed by traditional algebraic algorithms do not satisfy the clinical image quality demand.…”

A Model of Regularization Parameter Determination in Low-Dose X-Ray CT Reconstruction Based on Dictionary Learning

“…In this case, analytic-based algorithms like FDK [ 1 ], which are derived from a continuous imaging model and in need of dense sampled projections, are sensitive to insufficient projection data and arrive at a terrible result. However, algebraic algorithms like the simultaneous algebraic reconstruction technique (SART) [ 2 ] solved the problem better by transforming it to a series of linear equations.…”