Source-linked AI summary
GPU-based fast Monte Carlo simulation for radiotherapy dose calculation
Xun Jia, Xuejun Gu, Yan Jiang Graves, Michael Folkerts, Steve B. Jiang
TL;DR
Radiotherapy Monte Carlo dose calculation is accurate but slow, motivating a more efficient GPU implementation. The paper presents gDPM v2.0, which preserves DPM transport physics while separating photon and electron transport and adding clinical-plan support. It reports agreement with CPU results, 69.1 ~ 87.2 speedup, and 36.1~39.6 sec calculation for realistic IMRT or VMAT plans.
Problem
Monte Carlo dose calculation is accurate but requires long computation times, limiting efficiency for routine radiotherapy applications.
Method
gDPM v2.0 adapts DPM for GPU execution while preserving its particle transport physics and adding fluence-map, linac-geometry, and treatment-plan components.
Results
69.1 ~ 87.2 speedup was observed, while realistic IMRT or VMAT dose calculations completed in 36.1~39.6 sec with a single GPU.
Takeaways & Limitations
CPU and GPU dose differences were not statistically significant in over 96% of the high dose region and over 97% of the entire phantom region.
Takeaways & Limitations
Accurate linac source modeling is beyond the paper's scope; users must supply source-generation functions or a phase space file.
Abstract
from arXiv · showhide
Monte Carlo (MC) simulation is commonly considered to be the most accurate dose calculation method in radiotherapy. However, its efficiency still requires improvement for many routine clinical applications. In this paper, we present our recent progress towards the development a GPU-based MC dose calculation package, gDPM v2.0. It utilizes the parallel computation ability of a GPU to achieve high efficiency, while maintaining the same particle transport physics as in the original DPM code and hence the same level of simulation accuracy. In GPU computing, divergence of execution paths between threads can considerably reduce the efficiency. Since photons and electrons undergo different physics and hence attain different execution paths, we use a simulation scheme where photon transport and electron transport are separated to partially relieve the thread divergence issue. High performance random number generator and hardware linear interpolation are also utilized. We have also developed various components to handle fluence map and linac geometry, so that gDPM can be used to compute dose distributions for realistic IMRT or VMAT treatment plans. Our gDPM package is tested for its accuracy and efficiency in both phantoms and realistic patient cases. In all cases, the average relative uncertainties are less than 1%. A statistical t-test is performed and the dose difference between the CPU and the GPU results is found not statistically significant in over 96% of the high dose region and over 97% of the entire region. Speed up factors of 69.1 ~ 87.2 have been observed using an NVIDIA Tesla C2050 GPU card against a 2.27GHz Intel Xeon CPU processor. For realistic IMRT and VMAT plans, MC dose calculation can be completed with less than 1% standard deviation in 36.1~39.6 sec using gDPM.
1. Introduction
Monte Carlo dose calculation is highly accurate but computationally demanding, motivating GPU acceleration despite conflicts between Monte Carlo randomness and GPU execution. gDPM v2.0 addresses these conflicts while preserving DPM physics and supporting realistic treatment plans.
- Monte Carlo simulation provides accurate radiotherapy dose calculation but requires many particle histories and long computation times.
- GPU parallelization can improve efficiency, but Monte Carlo memory access and execution patterns limit achievable speed-up.
- gDPM v1.0 achieved only 5.0~6.6 times speedup by distributing particles across GPU threads.
- gDPM v2.0 preserves DPM particle transport physics while achieving 69.1 ~ 87.2 times higher efficiency than CPU DPM.
- Clinical-plan components enable Monte Carlo dose calculation for realistic IMRT and VMAT plans on a sub-minute time scale.
2.1 DPM Physics
DPM models coupled photon-electron transport using radiotherapy-appropriate approximations and established transport techniques. gDPM retains these physics while seeking computational efficiency improvements.
- DPM targets fast radiotherapy dose calculation by simulating coupled photon-electron transport with energy-range-specific approximations.
- Photon transport uses Woodcock tracking, while electron transport uses condensed history and continuous slowing down below preset energy thresholds.
- Positrons are treated as electrons and produce two photons at path termination to model annihilation.
- gDPM leaves DPM physics unchanged while pursuing an efficiency boost.
2.2 CUDA implementation
gDPM v2.0 restructures DPM for GPU execution using batched particle transport, GPU-oriented data structures, and separated photon and electron kernels. Additional implementation changes target numerical and sampling efficiency.
- CUDA implementation transfers voxel geometry, material properties, cross sections, and random seeds to GPU memory before batched simulation.
- Each GPU particle array contains particles of one type at a time, while separate secondary-particle stacks support parallel transport.
- Photon and electron transport are separated so the corresponding GPU kernel simulates one particle per thread in parallel.
- GPU implementation uses single-precision floating-point data instead of the original double-precision representation.
- gDPM v2.0 uses NVIDIA CURAND with the XORWOW algorithm for lightweight GPU random-number generation.
- Cross-section interpolation is modified from cubic splines to improve performance during simulation.
2.3 Other components in gDPM v2.0
gDPM v2.0 adds treatment-plan, fluence-map, energy-spectrum, and source interfaces for clinically realistic dose calculation. Its sampling approach approximates prescribed fluence increasingly well as particle count grows, while detailed linac source modeling remains user-supplied.
- Other components in gDPM v2.0: gDPM loads DICOM RT plans, patient CT and structures, fluence maps, and linac geometry for realistic IMRT or VMAT dose calculation.
- Fluence map: Fluence-map sampling represents beam angles and beamlet locations through indexed beamlets whose intensities define relative particle probabilities.
- Fluence map: Metropolis sampling avoids repeated fluence-map searches and their associated GPU memory accesses when selecting beamlets.
- Fluence map: The generated fluence-map error decreases monotonically and quickly as particle number increases.
- Energy spectrum: Photon energies must follow a spectrum, but different photon energies produce unequal simulation times and reduce GPU efficiency.
- Detailed source modeling: Accurate linac source modeling is outside the package scope, so users must provide source-generation functions or phase-space files.
3. Results
Phantom tests evaluated gDPM v2.0 for accuracy and computation time against the original CPU DPM implementation, using heterogeneous water–bone–water and water–lung–water geometries. The GPU results agreed with CPU results within statistical uncertainty while providing substantial speedups.
- Heterogeneous phantom tests: Phantom tests used water–bone–water and water–lung–water slab geometries with photon and electron beams.Depth-dose curves and lateral profiles were evaluated for both beam types in heterogeneous phantoms.
- Accuracy: The GPU and CPU dose distributions showed good agreement within statistical uncertainty in the heterogeneous phantom tests.The comparison included depth-dose curves and lateral dose profiles, with error bars representing two standard deviations.
- Accuracy: The average relative uncertainty was below 1% in all four phantom test cases.Uncertainty was averaged over the high-dose region, defined by local dose exceeding half the phantom’s maximum dose.
- Computation time: 69.1–87.2 was the observed GPU speed-up factor relative to the CPU implementation, including CPU–GPU data-transfer time in the GPU timing.The comparison used an NVIDIA Tesla C2050 GPU and the original DPM CPU implementation.
- Realistic treatment plans: Realistic-plan feasibility was assessed separately using one IMRT plan and two VMAT plans from the Varian Eclipse treatment-planning system.The plans included an eight-field noncoplanar head-and-neck IMRT case, a two-arc head-and-neck VMAT case, and a two-arc prostate VMAT case.
4. Discussion and Conclusions
gDPM v2.0 preserves the original DPM particle-transport physics while tailoring computation for GPUs and supporting clinically realistic treatment plans. It achieved close CPU–GPU agreement, large speedups, and sub-minute Monte Carlo dose calculations, although comparison with GPUMCD showed lower speed in the tested setup.
- Conclusions: gDPM v2.0 maintains the original DPM particle-transport physics while targeting GPU computation efficiency.The unchanged physics are intended to avoid degrading simulation accuracy.
- Efficiency: 69.1–87.2 was the observed speed-up factor for gDPM v2.0 against a 2.27 GHz Intel Xeon CPU processor.The reported comparison used an NVIDIA Tesla C2050 GPU.
- Clinical practicality: 36.1–39.6 sec was the reported time for Monte Carlo dose calculation of realistic IMRT or VMAT plans with a single GPU.The clinical-practicality components support dose calculation for realistic treatment plans.
- Future work: Further efficiency improvements were identified as desirable, including multi-GPU computation and variance-reduction techniques.The discussion states that shorter computation times would make GPU-based Monte Carlo dose calculation more clinically attractive.
- Comparison with related work: In the tested homogeneous-water comparison, gDPM v2.0 was approximately four times slower than GPUMCD.The reported absolute times were 0.47 and 0.90 sec for gDPM versus 0.12 and 0.27 sec for GPUMCD in the electron and photon cases, respectively.