Fast and Accurate Calculations for Proton Therapy

26
New Mexico Workshop, May 15, 2011 Pablo P. Yepes, Rice U. 1 Fast and Accurate Calculations for Proton Therapy Pablo Yepes 1,2 , John Eley 2 , Jessie Huang 1,2 , Dragan Mirkovic 2 , Wayne Newhauser 2 , Sharmalee Randeniya 2 , Uwe Titt 2 , Phil Taddei 2 1) T.W. Bonner Lab, Rice University 2) University of Texas, MD Anderson Cancer Center

Transcript of Fast and Accurate Calculations for Proton Therapy

Page 1: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Pablo P. Yepes, Rice U. 1

Fast and Accurate Calculations for Proton Therapy

Pablo Yepes 1,2, John Eley2, Jessie Huang 1,2, Dragan Mirkovic 2, Wayne Newhauser 2, Sharmalee Randeniya 2, Uwe Titt 2, Phil Taddei 2

1) T.W. Bonner Lab, Rice University2) University of Texas, MD Anderson Cancer Center

Page 2: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Pablo P. Yepes, Rice U. 2

Outline

FDC (track repeating algorithm) for dose calculations

Justification: MC vs Analytical Applications

GPU Implementation Applications, examples.

Page 3: Fast and Accurate Calculations for Proton Therapy

Why the Monte Carlo Approach

Pros: higher accuracy Information about off-field and

background (for example neutrons) radiation

Cons Very high computational demands

Page 4: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Pablo P. Yepes, Rice U. 4

Proton TherapyMonte Carlo versus Analytical

MC computational times are very high: >500 cpu hours for 1 beam treatment plan

Dose in Lung

AnalyticalMC (MCNPX)

Page 5: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Pablo P. Yepes, Rice U. 5

Proton TherapyMC versus Analytical (Con't)

Along beame

Page 6: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Pablo P. Yepes, Rice U. 6

Fast Dose Calculator (FDC)

A track repeating algorithm for proton dose calculations

Uses pre-calculated tracks in water to estimate dose in a variety of other materials by scaling path lengths and scattering angles.

Uses GEANT4 as engine to generate pre-calculated tracks. Same technique could be applied for the MC of your choice (MCNPX, FLUKA, etc).

• J. S. Li, B. Shahine, E. Fourkal and C.M. Ma, “A particle track-repeating algorithm for proton beam dose calculation”, Phys. Med. Biol., 50, 1001–10 (2005).

•P. Yepes, S. Randeniya, P. Taddei, W. Newhauser, A Track Repeating Algorithm for Fast Monte Carlo Dose Calculations of Proton Radiotherapy, Nuclear Technology 168, 334 (2009).

• P.Yepes P, S. Randeniya, P. J. Taddei and W. D. Newhauser, “Monte Carlo fast dose calculator for proton radiotherapy: application to a voxelized geometry representing a patient with prostate cancer”, Phys. Med. Biol., 54, N21-N28 (2009).

• Yepes, T. Brannan, J. Huang, D. Mirkovic, W.D. Newhauser, P.J. Taddei, U. Titt, “Application of a fast proton dosecalculation algorithm to a thorax geometry”, Rad. Meas. 45, 1367 (2010).

• P. Yepes, D. Mirkovic, P.J. Taddei, “A GPU implementation of a track-repeating algorithm for proton radiotherapy dose calculations”, Phys. Med. Biol. 55, 7107 (2010).

Page 7: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Pablo P. Yepes, Rice U. 7

Generate Pre-calculated tracks GEANT breaks trajectories in steps. For each step stores deposited energy, energy change, step length and direction. Simulate 250 MeV protons on a 660x660x750 mm3 water phantom. “Pick-up” the trajectory at whatever energy is needed. Average of 460 steps and 3.5Kbytes per proton Generate 10 million protons in water for database in ~15 min on 100 CPUs

Page 8: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Pablo P. Yepes, Rice U. 8

“Water history”

Re-trackinga “water history” inheterogeneousmedia

Page 9: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

FDC Scaling Parameters

Analytical Approach: Length scaling from Proton Stopping Power

(Bethe-Block Formula) Angle scaling from Multiple Scattering Theory.

Numerical Approach:

Obtain the scaling parameters by minimizing the difference in dose distributions in various homogeneous phantoms for various materials for FDC and GEANT4.

Page 10: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Pablo P. Yepes, Rice U. 10

Proton Simulation Setup

Aperture conforms the beam/field to the transverse shape of the tumor

Range compensator conforms the beam/field to the longitudinal shape of the tumor

Aperture

Range Compensator

Phantom

p

Page 11: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Pablo P. Yepes, Rice U. 11

GEANT4 versus FDCProtons Prostate

Realistic Beam and Shaping Elements (Aperture and Range Compensator

Page 12: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Pablo P. Yepes, Rice U. 12

Proton Prostate SimulationAccuracy/Timing

GEANT: 0.3 s/proton; FDC: 0.6 ms/proton

Accuracy relative to full Monte Carlo (GEANT) better than 2%

FDC Timing is ~ 500 faster than GEANT

Page 13: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Protons Lung• 140 MeV proton circular beam• Lung is a highly inhomogeneous medium, a tough test for FDC

FDCGEANT4

GEANT4-FDC

Page 14: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

FDC on GPUs

Page 15: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

FDC on GPUs (GFDC) Graphic Cards with multiple (GPU cores)

processors.

NVIDIA developed tools to allow running “regular” code on such cards.

Miniature computer cluster on a card for a fraction of the cost and maintenance.

Code divided on CPU and GPU sections

GPU code is regular (C and quasi C++) with some size and process limitations.

Each GPU core works on the re-tracking of one particular history.

Substantial reorganization of the code

GEFORCE GTX 295, NVIDIA

Page 16: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

GFDC Logic Flow

Page 17: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

GPU FDC for Protons

Circular Mono-energetic 121 MeV beam with 4 cm radius

Thorax phantom with 6 million voxels GFDC speed up of 75.5 with respect to FDC

CPU version

Protons/s

GEANT4 ~1

FDC 2,444

GFDC 184,525

Page 18: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Energy Deposited in Thorax for ProtonsAlong beam axis

Page 19: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Energy Deposited in Thorax for ProtonsCross Section: Axis Perpendicular to Beam

Page 20: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Lung Cross Section for ProtonsWhat about tails?

Page 21: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Accuracy

Page 22: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Statistical Errors/Timing

Batch

History-by-history

Page 23: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

FDC Website(Part of code from dicompyler: Adit Panchal, Roy Keyes)

Page 24: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Pablo P. Yepes, Rice U. 24

Conclusions Track-repeating algorithm for proton and photon

implemented on a GPU architecture. Computational times reduced by four orders of

magnitude compared to full Monte Carlo. GPU results reproduce full Monte Carlo within

2%. A system with multiple GPU cards should allow

real-time Monte-Carlo-accurate dose calculations for radio-therapy.

Page 25: Fast and Accurate Calculations for Proton Therapy

February 23, 2009 Pablo P. Yepes, Rice U. 25

Keep it simple!Keep it simple!

Page 26: Fast and Accurate Calculations for Proton Therapy

New Mexico Workshop, May 15, 2011

Pablo P. Yepes, Rice U. 26

Fast Dose Calculator in a Nutshell

Water to Material M scaling

For each step, length (L) and angle with respect to previous step (φ) are scaled.Length: LM= α (E) Lh2o

Scattering Angles: φM= β φh20

M=