Massively Parallel 3D Image Reconstruction

By Xiao Wang1; Amit Sabne2; Putt Sakdhnagool3; Sherman J. Kisner4; Charles Addison Bouman5; Sam Midkiff5

1. Harvard Medical School/Boston Children's Hospital, Boston, MA 2. Microsoft, Redmond, WA 3. Purdue University, West Lafayette, IN 4. High Performance Imaging LLC, West lafayette, IN 5. Electrical and Computer Engineering, Purdue University, West Lafayette, IN

Published on

Abstract

ACM Gordon Bell Finalists

Computed Tomography (CT) image reconstruction is an important technique used in a wide range of applications. Among reconstruction methods, Model-Based Iterative Reconstruction (MBIR) generally produces higher quality images. However, the irregular data access pattern, the difficulty of effective parallelization and slow algorithmic convergence have made MBIR impractical for many applications. This paper presents a new algorithm for MBIR, Non-Uniform Parallel Super-Voxel (NU-PSV), that regularizes the data access pattern, enables massive parallelism and ensures fast convergence. We compare the NU-PSV algorithm with two state-of-the-art implementations on a 69632-core distributed system. Results indicate that the NU-PSV algorithm has an average speedup of 1665 compared to the fastest state-of-the-art implementations.

Download Paper (PDF)

Bio

Xiao Wang Xiao Wang received B.A. degrees in Mathematics and Computer Science with honor from Saint John’s University, MN, in 2012, and M.S. degree in electrical and compute engineering from Purdue University, West Lafayette, IN, in 2016. In 2017, He received a PhD degree in electrical and computer engineering from Purdue University, under the supervision of Professor Charles Bouman and Samuel Midkiff. Currently, he works as a postdoctoral research fellow both at Harvard Medical School and Boston Children's Hospital.

Xiao Wang’s research work focuses on applying high performance computing to imaging problems, especially image processing on CT and MRI. He and his advisors are selected to be the 2017 ACM Gordon Bell Prize finalist for their research work on CT image reconstructions.

Sponsored by

Cite this work

Researchers should cite this work as follows:

  • Xiao Wang, Amit Sabne, Putt Sakdhnagool, Sherman J. Kisner, Charles A. Bouman, and Samuel P.Midkiff, Massively Parallel 3D Image Reconstruction, Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, SC '17, Denver, Colorado, (2017), doi: 10.1145/3126908.3126911

  • Xiao Wang, Amit Sabne, Putt Sakdhnagool, Sherman J. Kisner, Charles Addison Bouman, Sam Midkiff (2018), "Massively Parallel 3D Image Reconstruction," https://nanohub.org/resources/27760.

    BibTex | EndNote

Time

Location

SC17, Colorado Convention Center, Denver, CO

Massively Parallel 3D Image Reconstruction
  • Massively Parallel 3D Image Reconstruction* 1. Massively Parallel 3D Image Re… 0
    00:00/00:00
  • Tomographic Reconstruction 2. Tomographic Reconstruction 34.601267934601267
    00:00/00:00
  • Tomographic Reconstruction 3. Tomographic Reconstruction 80.980980980980988
    00:00/00:00
  • Model Based Iterative Reconstruction (MBIR) 4. Model Based Iterative Reconstr… 108.74207540874208
    00:00/00:00
  • Model Based Iterative Reconstruction (MBIR) 5. Model Based Iterative Reconstr… 152.78611945278612
    00:00/00:00
  • Irregular Memory Layout 6. Irregular Memory Layout 246.21287954621289
    00:00/00:00
  • Single-Core Performance Bottlenecks 7. Single-Core Performance Bottle… 323.82382382382383
    00:00/00:00
  • Parallel Computation Bottlenecks 8. Parallel Computation Bottlenec… 348.64864864864865
    00:00/00:00
  • Overview of Results 9. Overview of Results 448.34834834834834
    00:00/00:00
  • Efficient Computations of MBIR 10. Efficient Computations of MBIR 509.00900900900905
    00:00/00:00
  • Super-Voxel: Increasing Cache Utilization 11. Super-Voxel: Increasing Cache … 523.79045712379047
    00:00/00:00
  • Three-Dimensional Super-Voxel (3DSV) 12. Three-Dimensional Super-Voxel … 573.20653987320657
    00:00/00:00
  • Overview: 3 Levels of Parallelization 13. Overview: 3 Levels of Parallel… 650.38371705038378
    00:00/00:00
  • 3D Super-Voxel Buffer 14. 3D Super-Voxel Buffer 712.278945612279
    00:00/00:00
  • Block-Transposed Buffer (BTB) 15. Block-Transposed Buffer (BTB) 753.85385385385393
    00:00/00:00
  • Shared Memory Parallelism 16. Shared Memory Parallelism 835.23523523523522
    00:00/00:00
  • Multi-Node Parallelism 17. Multi-Node Parallelism 905.10510510510517
    00:00/00:00
  • Benchmark Dataset 18. Benchmark Dataset 938.00467133800475
    00:00/00:00
  • Iron Hydroxide AFRL Dataset Speedup 19. Iron Hydroxide AFRL Dataset Sp… 961.79512846179512
    00:00/00:00
  • Berkeley Biological Dataset 20. Berkeley Biological Dataset 1086.0193526860194
    00:00/00:00
  • Berkeley Biological Dataset Speedup 21. Berkeley Biological Dataset Sp… 1112.7127127127128
    00:00/00:00
  • SIMD Speedup 22. SIMD Speedup 1157.4908241574908
    00:00/00:00
  • Strong Scaling Efficiency 23. Strong Scaling Efficiency 1193.8271604938273
    00:00/00:00
  • Conclusions 24. Conclusions 1215.4154154154155
    00:00/00:00