SU‐E‐J‐60: GPU Accelerated DRR Computation with Scatter

F. Vidal, M. Folkerts, N. Freud, S. Jiang

Research output: Contribution to journalArticle

Abstract

Purpose: We propose a fast software library implemented on graphics processing unit (GPU) to compute digitally reconstructed radiographs (DRRs). It takes into account first order Compton scattering. Methods: The simulation is based on the evaluation of the Beer‐Lambert law and of the Klein‐Nishina equation. The algorithm is fully determinist and has been fully implemented on GPU to achieve clinically acceptable efficiency. A full resolution simulation is performed for primary radiation. A much lower image resolution is used for Compton scattering as it adds a low frequency pattern over the projection image. Each voxel of the CT dataset is considered as a secondary source. The number of photons that reach each voxel is evaluated. Then, for each secondary source, a projection image is computed and integrated in the final image. The photon energy between each secondary source and each pixel is also computed. An interlaced sampling mode is also proposed to further reduce the computation time without sacrificing numerical accuracy. Finally, the speed and accuracy are assessed.Results: We show that the computations can be fully implemented on the GPU with an original under‐sampling method to produce clinically acceptable results. For example, a simulation can be achieved in less than 7 seconds whilst the maximum relative error remains below 5% and the average relative error below 1.4%. At full resolution, a speed‐up by factor ∼12X is achieved for the GPU implementation with our interlaced‐mode by comparison with our multi‐threaded CPU implementation using 8 threads in parallel Conclusions: DRR calculation with scatter is computationally intensive. The use of GPU can achieve clinically acceptable efficiency. A Compton fluence map can be computed in a few seconds using under‐sampling, whilst keeping numerical inaccuracies relatively low. This work can be used for CBCT reconstruction to reduce scatter artifacts.

Original languageEnglish (US)
Pages (from-to)3455-3456
Number of pages2
JournalMedical Physics
Volume38
Issue number6
DOIs
StatePublished - 2011

Fingerprint

Photons
Artifacts
Software
Radiation
Datasets

ASJC Scopus subject areas

  • Biophysics
  • Radiology Nuclear Medicine and imaging

Cite this

SU‐E‐J‐60 : GPU Accelerated DRR Computation with Scatter. / Vidal, F.; Folkerts, M.; Freud, N.; Jiang, S.

In: Medical Physics, Vol. 38, No. 6, 2011, p. 3455-3456.

Research output: Contribution to journalArticle

Vidal, F. ; Folkerts, M. ; Freud, N. ; Jiang, S. / SU‐E‐J‐60 : GPU Accelerated DRR Computation with Scatter. In: Medical Physics. 2011 ; Vol. 38, No. 6. pp. 3455-3456.
@article{04ee9744a25549efa6a01a3db0f4984d,
title = "SU‐E‐J‐60: GPU Accelerated DRR Computation with Scatter",
abstract = "Purpose: We propose a fast software library implemented on graphics processing unit (GPU) to compute digitally reconstructed radiographs (DRRs). It takes into account first order Compton scattering. Methods: The simulation is based on the evaluation of the Beer‐Lambert law and of the Klein‐Nishina equation. The algorithm is fully determinist and has been fully implemented on GPU to achieve clinically acceptable efficiency. A full resolution simulation is performed for primary radiation. A much lower image resolution is used for Compton scattering as it adds a low frequency pattern over the projection image. Each voxel of the CT dataset is considered as a secondary source. The number of photons that reach each voxel is evaluated. Then, for each secondary source, a projection image is computed and integrated in the final image. The photon energy between each secondary source and each pixel is also computed. An interlaced sampling mode is also proposed to further reduce the computation time without sacrificing numerical accuracy. Finally, the speed and accuracy are assessed.Results: We show that the computations can be fully implemented on the GPU with an original under‐sampling method to produce clinically acceptable results. For example, a simulation can be achieved in less than 7 seconds whilst the maximum relative error remains below 5{\%} and the average relative error below 1.4{\%}. At full resolution, a speed‐up by factor ∼12X is achieved for the GPU implementation with our interlaced‐mode by comparison with our multi‐threaded CPU implementation using 8 threads in parallel Conclusions: DRR calculation with scatter is computationally intensive. The use of GPU can achieve clinically acceptable efficiency. A Compton fluence map can be computed in a few seconds using under‐sampling, whilst keeping numerical inaccuracies relatively low. This work can be used for CBCT reconstruction to reduce scatter artifacts.",
author = "F. Vidal and M. Folkerts and N. Freud and S. Jiang",
year = "2011",
doi = "10.1118/1.3611828",
language = "English (US)",
volume = "38",
pages = "3455--3456",
journal = "Medical Physics",
issn = "0094-2405",
publisher = "AAPM - American Association of Physicists in Medicine",
number = "6",

}

TY - JOUR

T1 - SU‐E‐J‐60

T2 - GPU Accelerated DRR Computation with Scatter

AU - Vidal, F.

AU - Folkerts, M.

AU - Freud, N.

AU - Jiang, S.

PY - 2011

Y1 - 2011

N2 - Purpose: We propose a fast software library implemented on graphics processing unit (GPU) to compute digitally reconstructed radiographs (DRRs). It takes into account first order Compton scattering. Methods: The simulation is based on the evaluation of the Beer‐Lambert law and of the Klein‐Nishina equation. The algorithm is fully determinist and has been fully implemented on GPU to achieve clinically acceptable efficiency. A full resolution simulation is performed for primary radiation. A much lower image resolution is used for Compton scattering as it adds a low frequency pattern over the projection image. Each voxel of the CT dataset is considered as a secondary source. The number of photons that reach each voxel is evaluated. Then, for each secondary source, a projection image is computed and integrated in the final image. The photon energy between each secondary source and each pixel is also computed. An interlaced sampling mode is also proposed to further reduce the computation time without sacrificing numerical accuracy. Finally, the speed and accuracy are assessed.Results: We show that the computations can be fully implemented on the GPU with an original under‐sampling method to produce clinically acceptable results. For example, a simulation can be achieved in less than 7 seconds whilst the maximum relative error remains below 5% and the average relative error below 1.4%. At full resolution, a speed‐up by factor ∼12X is achieved for the GPU implementation with our interlaced‐mode by comparison with our multi‐threaded CPU implementation using 8 threads in parallel Conclusions: DRR calculation with scatter is computationally intensive. The use of GPU can achieve clinically acceptable efficiency. A Compton fluence map can be computed in a few seconds using under‐sampling, whilst keeping numerical inaccuracies relatively low. This work can be used for CBCT reconstruction to reduce scatter artifacts.

AB - Purpose: We propose a fast software library implemented on graphics processing unit (GPU) to compute digitally reconstructed radiographs (DRRs). It takes into account first order Compton scattering. Methods: The simulation is based on the evaluation of the Beer‐Lambert law and of the Klein‐Nishina equation. The algorithm is fully determinist and has been fully implemented on GPU to achieve clinically acceptable efficiency. A full resolution simulation is performed for primary radiation. A much lower image resolution is used for Compton scattering as it adds a low frequency pattern over the projection image. Each voxel of the CT dataset is considered as a secondary source. The number of photons that reach each voxel is evaluated. Then, for each secondary source, a projection image is computed and integrated in the final image. The photon energy between each secondary source and each pixel is also computed. An interlaced sampling mode is also proposed to further reduce the computation time without sacrificing numerical accuracy. Finally, the speed and accuracy are assessed.Results: We show that the computations can be fully implemented on the GPU with an original under‐sampling method to produce clinically acceptable results. For example, a simulation can be achieved in less than 7 seconds whilst the maximum relative error remains below 5% and the average relative error below 1.4%. At full resolution, a speed‐up by factor ∼12X is achieved for the GPU implementation with our interlaced‐mode by comparison with our multi‐threaded CPU implementation using 8 threads in parallel Conclusions: DRR calculation with scatter is computationally intensive. The use of GPU can achieve clinically acceptable efficiency. A Compton fluence map can be computed in a few seconds using under‐sampling, whilst keeping numerical inaccuracies relatively low. This work can be used for CBCT reconstruction to reduce scatter artifacts.

UR - http://www.scopus.com/inward/record.url?scp=85024798369&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85024798369&partnerID=8YFLogxK

U2 - 10.1118/1.3611828

DO - 10.1118/1.3611828

M3 - Article

AN - SCOPUS:85024798369

VL - 38

SP - 3455

EP - 3456

JO - Medical Physics

JF - Medical Physics

SN - 0094-2405

IS - 6

ER -