Encrypted login | home

Program Information

Fast Dose Calculation for Magnetic-Resonance Imaging-Guided Radiation Therapy (MRIgRT) Using GPU-Based Monte Carlo Code ARCHER

no image available
T Liu

T Liu1*, H Lin1 , L Yang2 , H Liu2 , Z Wang2,3 , X Pei2 , Z Chen2 , X Xu1,2 , (1) Rensselaer Polytechnic Institute, Troy, New York, (2) University of Science and Technology of China, Hefei, Anhui, (3) The First Affiliated Hospital of Anhui Medical University, Hefei, Anhui


TU-D-205-5 (Tuesday, August 1, 2017) 11:00 AM - 12:15 PM Room: 205

Purpose: The development and deployment of magnetic-resonance imaging-guided radiation therapy (MRIgRT) systems have gained momentum in recent years. Additional complexity in dose verification is introduced due to the fact that magnetic field changes the trajectory of charged particles and the subsequent dose distribution. In this study, we extended our GPU-accelerated Monte Carlo code ARCHER to MRIgRT dose calculation with high performance and accuracy.

Methods: The fast transport method in a uniform magnetic field is based on the first order approximation to particle velocity as is outlined in Salvat 2015 (Penelope). ARCHER was verified against MCNP-6.1 and Geant4-10.02 using a 10x10x10 cm3 water phantom, 20 MeV electron and positron beams in a 1.5 T magnetic field. In MCNP the detailed transport algorithm originating from the MARS code was used, while in Geant4 the efficient "Nystrom RK4" algorithm was selected. Dose calculation was performed for a clinical breast cancer treatment using a Truebeam 6MV system.

Results: ARCHER was in good agreement with Geant4, achieving 99.999% and 99.983% pass rate in the 3\%/3mm gamma test for electron and positron cases, respectively. ARCHER on a Titan X GPU showed superior performance to the multi-threaded MCNP and Geant4 on a high-end 14-core (28 threads) E5 2697 v3 CPU. Simulation of 10 million primary particles took 7 seconds on ARCHER, 2 days on MCNP and 1.8 hours on Geant4. For the clinical case, ARCHER took 32, 48, and 46 seconds to complete the simulation on a Titan X, K40 and Quadro M6000 GPU, respectively. The significant performance improvement is mainly attributed to the appropriately simplified physics and efficient, GPU-specific micro-optimizations in ARCHER.

Conclusion: A fast and accurate method was implemented in ARCHER, allowing fast dose calculation for MRIgRT.

Contact Email: