monte carlo dose calculations for proton therapy on gpu
TRANSCRIPT
DEPT OF RADIATION ONCOLOGY
Monte Carlo dose calculations for
proton therapy on GPU
Xun Jia1, Jan Schuemann2, Drosoula Giantsoudi2, Harald Paganetti2, Steve Jiang1
1University of Texas Southwestern Medical Center 2Massachusetts General Hospital, Harvard Medical School
DEPT OF RADIATION ONCOLOGY
CAMPUS PROPER
NORTH CAMPUS
WEST CAMPUS
CLINICAL CAMPUS
SOUTH CAMPUS
PARKLAND HOSPITAL
UT Southwestern
Texas Center for Advanced
Radiation Therapy (TCART)
Current Facility
(w/ 10 Tx machines)
DEPT OF RADIATION ONCOLOGY
SCHEME A – MASSING
Phase I: Proton Center $286M, 4 gantries + 1 fixed beam Under construction
Texas Center for Advanced Radiation Therapy (TCART)
Phase II: Conventional Center $62-92M, 8+5 linac vaults Under designing
Phase III: Carbon Ion Center ~$200M Under planning
DEPT OF RADIATION ONCOLOGY
gPMC project
Monte Carlo dose calculations for proton therapy
Objective: speed up full MC dose using Graphics
Processing Units (GPU)
Develop an appropriate physics model to achieve sufficient
accuracy for proton therapy
Design GPU-friendly implementations to achieve high
computational efficiency
DEPT OF RADIATION ONCOLOGY
Proton transport Proton transport physics
Combines the physics from various literatures Kawrakow, Med Phys, 27, 485(2000), Fippel et. al., Med Phys, 31, 2263 (2004),
Penelope manual (2009), Geant4 physics manual (2011)
Energy range [0.5, 350.0] MeV
Class II condensed history simulation scheme with continuous
slowing down approximation
Multiple scattering and energy straggling
CSDA for secondary electron transport
Nuclear interaction is handled by an empirical strategy Fippel et. al., Med Phys, 31, 2263(2004)
Support 25 materials that were defined in TOPAS/Geant4
Perl et. al. Submitted to Med Phys(2012)
Voxelized patient geometry
CT to material conversion
DEPT OF RADIATION ONCOLOGY
GPU A specialized accelerator (personal super-computer)
Implementation
Nvidia CUDA environment
Optimized parallelization scheme to reduce GPU-thread divergence
Multi-dose counter to reduce memory writing conflict
Hardware supported interpolation
High-efficiency random number generator
4-16 cores >1000 cores
Jia, et. al., PMB 57, 7783 (2012)
DEPT OF RADIATION ONCOLOGY
Homogeneous phantoms Pure water
200 MeV
Other homogeneous phantoms
100, 200 MeV
Pγ = 99.9%
(2mm/2%)
Pγ > 98.7%
(2mm/2%)
DEPT OF RADIATION ONCOLOGY
Inhomogeneous phantoms
Materials
Water (0 HU)
Bone insert (500 HU)
Air insert (-1000 HU)
air
bone
water
Asquarebeam
x
z y
(a)
0 5 10 15 20 25 30
0
0.2
0.4
0.6
0.8
1
1.2
0 4 8 12
z (cm)
0.1
0.6
1.1
D (MeV/g)
TOPAS/Geant4, bonegPMC, boneTOPAS/Geant4, airgPMC, air
(b) (c)
bone air
0 5 10 15 20 25 30
0
0.2
0.4
0.6
0.8
1
1.2(b)
-4 -2 0 2 4
x (cm)
0.1
0.6
1.1
D (MeV/g)
TOPAS/Geant4, 6cmTOPAS/Geant4, 12cmgPMC, 6cmgPMC, 12cm
(c)
bone air
lung
bone
water
Asquarebeam
x
z y
(a)
0 5 10 15 20 25 30
0
0.2
0.4
0.6
0.8
1
1.2
0 4 8 12
z (cm)
0
1
2
D (MeV/g)
TOPAS/Geant4, boneTOPAS/Geant4, lunggPMC, bonegDPM, lung
(b)
-4 -2 0 2 4
x (cm)
0
1
2
D (MeV/g)
TOPAS/Geant4, 6.95cmTOPAS/Geant4, 10.85cmgPMC, 6.95cmgPMC, 10.85cm
(c)
bone lung
0 5 10 15 20 25 30
0
0.2
0.4
0.6
0.8
1
1.2
0 4 8 12
z (cm)
0
1
2
D (MeV/g)
TOPAS/Geant4, boneTOPAS/Geant4, lunggPMC, bonegDPM, lung
(b)
-4 -2 0 2 4
x (cm)
0
1
2
D (MeV/g)
TOPAS/Geant4, 6.95cmTOPAS/Geant4, 10.85cmgPMC, 6.95cmgPMC, 10.85cm
(c)
bone lung
Pγ = 99.9%
(2mm/2%)
DEPT OF RADIATION ONCOLOGY
Patient case
gPMC TOPAS/GEANT4
γ-index distribution
(2%/2mm)
CTV: Pγ = 99.75%
GTV: Pγ = 99.90%
Whole patient:
Pγ = 99.94%
DEPT OF RADIATION ONCOLOGY
Efficiency
GPU: NVIDIA Tesla C2050
CPU computation time is 2~80 CPU hrs
Phantom Source
Energy (MeV)
<σ/D>
(%)
T
(sec)
Water 100 0.9 6.76
200 1.1 21.14
Bone 100 0.9 6.68
200 1.1 20.98
Tissue 100 0.9 7.08
200 1.1 22.29
Inhomogeneous phantom 100 0.9 9.44
Patient # Beam # Npart
(M)
Tload
(sec)
Tsim
(sec)
Pγ
(%)
1
1 10.8 15.5 16.2 97.5
2 10.7 15.2 13.3 98.0
3 7.6 11.2 10.8 97.4
4 8.4 12.5 12.6 97.7
5 5.6 9.2 8.6 97.4
2
1 6.8 10.6 5.5 99.2
2 7.1 10.5 5.7 99.4
3 7.3 10.7 5.8 99.6
DEPT OF RADIATION ONCOLOGY
Summary
gPMC: GPU-based MC dose calculation package for proton
therapy
The accuracy is validated against TOPAS/Geant4
Dose calculation time is 6~22 sec
http://www.forbes.com/sites/netapp/2014/06/02/video
-game-cure-cancer/
This work was featured by Forbes Magazine
on June 2nd, 2014
The 5th Short Course in GPU Programming on 9/3-4
The code can be shared for research purpose