TH-CD-303-02: A GPU-Based Iterative Image Reconstruction Solver with 4D Regularization for Low-Dose Helical 4DCT

Authors

  • Guo M,

    1. Shanghai Jiao Tong University, Shanghai, Shanghai
    2. Stanford University, Palo Alto, CA
    3. Stanford Univ School of Medicine, Stanford, CA
    4. Shanghai Jiao Tong University, Shanghai, Shanghai
    Search for more papers by this author
  • Li R,

    1. Shanghai Jiao Tong University, Shanghai, Shanghai
    2. Stanford University, Palo Alto, CA
    3. Stanford Univ School of Medicine, Stanford, CA
    4. Shanghai Jiao Tong University, Shanghai, Shanghai
    Search for more papers by this author
  • Xing L,

    1. Shanghai Jiao Tong University, Shanghai, Shanghai
    2. Stanford University, Palo Alto, CA
    3. Stanford Univ School of Medicine, Stanford, CA
    4. Shanghai Jiao Tong University, Shanghai, Shanghai
    Search for more papers by this author
  • Gao H

    1. Shanghai Jiao Tong University, Shanghai, Shanghai
    2. Stanford University, Palo Alto, CA
    3. Stanford Univ School of Medicine, Stanford, CA
    4. Shanghai Jiao Tong University, Shanghai, Shanghai
    Search for more papers by this author

Abstract

Purpose:

4DCT is an important technique for motion management in radiotherapy, but involves a significant amount of dose to patients. Low-dose helical 4DCT scans can be done by (a) lowering tube current and pulse duration, (b) increasing helical pitch, and (c) reducing data sampling. Conventional FDK methods may fail for these low-dose scan protocols due to noisy or incomplete data. We aim to develop an accurate and fast image reconstruction solver that enables dose reduction for helical 4DCT.

Methods:

For improved accuracy, the developed solver is based on a novel 4D regularization method for CT image reconstruction, which promotes both the smoothness in spatial variables for each phase and the similarity in temporal variables between phases, in contrast with the 3D regularization method with spatial regularization only. Both 3D and 4D methods here are based on total variation. The formulated optimization problem is solved by alternating direction method of multipliers (ADMM), which is chosen for its effectiveness for solving the convex problem, and more importantly its suitability for distributed computing.For improved speed, the developed solver utilizes an efficient parallel algorithm for X-ray transform and its adjoint with optimized computational cost per parallel thread O (1) that takes flying focal spots into account, and multi-GPU platforms for solving such a massive parallel computing problem in helical 4DCT.

Results:

The solver was validated using both simulations (4D NCAT phantoms) and experimental sinogram data (Siemens, Somatom 64-slices) with low-dose, high-pitch and sparse-view helical 4DCT scans. Both the reconstructed images and quantitative results demonstrate that the 4D method improved image quality compared with the 3D method, which is in turn better than FDK.

Conclusion:

We have developed a fast, accurate solver that synergizes state-of-art iterative reconstruction methods and parallel computing techniques that can potentially enable low-dose, high-pitch and sparse-view helical 4DCT scans.

Minghao Guo and Hao Gao were partially supported by the NSFC (#11405105), the 973 Program (#2015CB856000) and the Shanghai Pujiang Talent Program (#14PJ1404500).

Ancillary