SU‐DD‐A4‐03: Cone Beam Computed Tomography Reconstruction on Graphics Processing Units: Flat3D Texture Techniques

J. Kim, L. Ren, JianYue Jin, H. Zhong, I. J. Chetty

Research output: Contribution to journalArticle

Abstract

Purpose: Graphics processing units (GPUs) can shorten cone‐beam volume reconstruction time significantly in comparison to CPUs. We evaluated Flat3D texture techniques to further enhance calculation efficiency. Methods: Typical reconstruction involves processing of a set of 2D projection images, and filtered backprojection of them onto volume slices. We implemented three different back‐projection algorithms on GPUs. The first method, BP_tex2d, represents the volume as a set of 2D textures and backprojection onto the texture occurs one slice at a time. The second method (BP_flat3d) represents the volume using a Flat3D texture. Flat3D texture is a 2D texture that maps volume slices onto 2D texture surface side‐by‐side. Therefore, for a given projection image, the back‐projection is accomplished in a single rendering pass. The third approach (BP_flat3d_multitex) back‐projects multiple projection images (16 here) onto a Flat3D texture in a single rendering pass using multi‐texture support. We reconstructed a 3D Shepp‐Logan phantom (256×256×256 pixel 3 ) from a set of 360 projection images (512×384 pixel 2 ) using each method. Back‐projection time was estimated by subtracting the image load and filter time from the total reconstruction time. Ten measurements were performed for each method and the mean time determined. Host computer: 3.6 GHz Intel Pentium 4 CPU with 2 GB memory. GPU: Nvidia_GeForce_210 (512 MB)+PCIe. Results: The mean back projection times of BP_tex2d, BP_flat3d, and BP_flat3d_multitex were 51, 16, and 8 seconds respectively. The image loading and filtering time was 24 seconds. The reconstructed volumes had identical pixel values. Conclusion; Flat3D texture approach improved the back‐projection time by a factor of 3.2, compared to regular 2D texture based method. The utilization of multi‐texture improved the speedup by a factor of 2 (thus, 6.4 total). High quality iterative reconstruction may be possible in the clinical setting given such improvements on multiple GPUs.

Original languageEnglish (US)
Number of pages1
JournalMedical Physics
Volume37
Issue number6
DOIs
StatePublished - Jan 1 2010

Fingerprint

Cone-Beam Computed Tomography
Efficiency

ASJC Scopus subject areas

  • Biophysics
  • Radiology Nuclear Medicine and imaging

Cite this

SU‐DD‐A4‐03 : Cone Beam Computed Tomography Reconstruction on Graphics Processing Units: Flat3D Texture Techniques. / Kim, J.; Ren, L.; Jin, JianYue; Zhong, H.; Chetty, I. J.

In: Medical Physics, Vol. 37, No. 6, 01.01.2010.

Research output: Contribution to journalArticle

Kim, J. ; Ren, L. ; Jin, JianYue ; Zhong, H. ; Chetty, I. J. / SU‐DD‐A4‐03 : Cone Beam Computed Tomography Reconstruction on Graphics Processing Units: Flat3D Texture Techniques. In: Medical Physics. 2010 ; Vol. 37, No. 6.
@article{cb6ec0a9905747489be92aee6547d2cd,
title = "SU‐DD‐A4‐03: Cone Beam Computed Tomography Reconstruction on Graphics Processing Units: Flat3D Texture Techniques",
abstract = "Purpose: Graphics processing units (GPUs) can shorten cone‐beam volume reconstruction time significantly in comparison to CPUs. We evaluated Flat3D texture techniques to further enhance calculation efficiency. Methods: Typical reconstruction involves processing of a set of 2D projection images, and filtered backprojection of them onto volume slices. We implemented three different back‐projection algorithms on GPUs. The first method, BP_tex2d, represents the volume as a set of 2D textures and backprojection onto the texture occurs one slice at a time. The second method (BP_flat3d) represents the volume using a Flat3D texture. Flat3D texture is a 2D texture that maps volume slices onto 2D texture surface side‐by‐side. Therefore, for a given projection image, the back‐projection is accomplished in a single rendering pass. The third approach (BP_flat3d_multitex) back‐projects multiple projection images (16 here) onto a Flat3D texture in a single rendering pass using multi‐texture support. We reconstructed a 3D Shepp‐Logan phantom (256×256×256 pixel 3 ) from a set of 360 projection images (512×384 pixel 2 ) using each method. Back‐projection time was estimated by subtracting the image load and filter time from the total reconstruction time. Ten measurements were performed for each method and the mean time determined. Host computer: 3.6 GHz Intel Pentium 4 CPU with 2 GB memory. GPU: Nvidia_GeForce_210 (512 MB)+PCIe. Results: The mean back projection times of BP_tex2d, BP_flat3d, and BP_flat3d_multitex were 51, 16, and 8 seconds respectively. The image loading and filtering time was 24 seconds. The reconstructed volumes had identical pixel values. Conclusion; Flat3D texture approach improved the back‐projection time by a factor of 3.2, compared to regular 2D texture based method. The utilization of multi‐texture improved the speedup by a factor of 2 (thus, 6.4 total). High quality iterative reconstruction may be possible in the clinical setting given such improvements on multiple GPUs.",
author = "J. Kim and L. Ren and JianYue Jin and H. Zhong and Chetty, {I. J.}",
year = "2010",
month = "1",
day = "1",
doi = "10.1118/1.3468001",
language = "English (US)",
volume = "37",
journal = "Medical Physics",
issn = "0094-2405",
publisher = "AAPM - American Association of Physicists in Medicine",
number = "6",

}

TY - JOUR

T1 - SU‐DD‐A4‐03

T2 - Cone Beam Computed Tomography Reconstruction on Graphics Processing Units: Flat3D Texture Techniques

AU - Kim, J.

AU - Ren, L.

AU - Jin, JianYue

AU - Zhong, H.

AU - Chetty, I. J.

PY - 2010/1/1

Y1 - 2010/1/1

N2 - Purpose: Graphics processing units (GPUs) can shorten cone‐beam volume reconstruction time significantly in comparison to CPUs. We evaluated Flat3D texture techniques to further enhance calculation efficiency. Methods: Typical reconstruction involves processing of a set of 2D projection images, and filtered backprojection of them onto volume slices. We implemented three different back‐projection algorithms on GPUs. The first method, BP_tex2d, represents the volume as a set of 2D textures and backprojection onto the texture occurs one slice at a time. The second method (BP_flat3d) represents the volume using a Flat3D texture. Flat3D texture is a 2D texture that maps volume slices onto 2D texture surface side‐by‐side. Therefore, for a given projection image, the back‐projection is accomplished in a single rendering pass. The third approach (BP_flat3d_multitex) back‐projects multiple projection images (16 here) onto a Flat3D texture in a single rendering pass using multi‐texture support. We reconstructed a 3D Shepp‐Logan phantom (256×256×256 pixel 3 ) from a set of 360 projection images (512×384 pixel 2 ) using each method. Back‐projection time was estimated by subtracting the image load and filter time from the total reconstruction time. Ten measurements were performed for each method and the mean time determined. Host computer: 3.6 GHz Intel Pentium 4 CPU with 2 GB memory. GPU: Nvidia_GeForce_210 (512 MB)+PCIe. Results: The mean back projection times of BP_tex2d, BP_flat3d, and BP_flat3d_multitex were 51, 16, and 8 seconds respectively. The image loading and filtering time was 24 seconds. The reconstructed volumes had identical pixel values. Conclusion; Flat3D texture approach improved the back‐projection time by a factor of 3.2, compared to regular 2D texture based method. The utilization of multi‐texture improved the speedup by a factor of 2 (thus, 6.4 total). High quality iterative reconstruction may be possible in the clinical setting given such improvements on multiple GPUs.

AB - Purpose: Graphics processing units (GPUs) can shorten cone‐beam volume reconstruction time significantly in comparison to CPUs. We evaluated Flat3D texture techniques to further enhance calculation efficiency. Methods: Typical reconstruction involves processing of a set of 2D projection images, and filtered backprojection of them onto volume slices. We implemented three different back‐projection algorithms on GPUs. The first method, BP_tex2d, represents the volume as a set of 2D textures and backprojection onto the texture occurs one slice at a time. The second method (BP_flat3d) represents the volume using a Flat3D texture. Flat3D texture is a 2D texture that maps volume slices onto 2D texture surface side‐by‐side. Therefore, for a given projection image, the back‐projection is accomplished in a single rendering pass. The third approach (BP_flat3d_multitex) back‐projects multiple projection images (16 here) onto a Flat3D texture in a single rendering pass using multi‐texture support. We reconstructed a 3D Shepp‐Logan phantom (256×256×256 pixel 3 ) from a set of 360 projection images (512×384 pixel 2 ) using each method. Back‐projection time was estimated by subtracting the image load and filter time from the total reconstruction time. Ten measurements were performed for each method and the mean time determined. Host computer: 3.6 GHz Intel Pentium 4 CPU with 2 GB memory. GPU: Nvidia_GeForce_210 (512 MB)+PCIe. Results: The mean back projection times of BP_tex2d, BP_flat3d, and BP_flat3d_multitex were 51, 16, and 8 seconds respectively. The image loading and filtering time was 24 seconds. The reconstructed volumes had identical pixel values. Conclusion; Flat3D texture approach improved the back‐projection time by a factor of 3.2, compared to regular 2D texture based method. The utilization of multi‐texture improved the speedup by a factor of 2 (thus, 6.4 total). High quality iterative reconstruction may be possible in the clinical setting given such improvements on multiple GPUs.

UR - http://www.scopus.com/inward/record.url?scp=85024781950&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85024781950&partnerID=8YFLogxK

U2 - 10.1118/1.3468001

DO - 10.1118/1.3468001

M3 - Article

AN - SCOPUS:85024781950

VL - 37

JO - Medical Physics

JF - Medical Physics

SN - 0094-2405

IS - 6

ER -