Unencrypted login | home

Program Information

Affordable Supercomputer-Based Monte Carlo CT Dose Calculations: A Hardware Comparison Between Nvidia M2090 GPU and Intel Xeon Phi 5110p Coprocessor

no image available
T Liu

T Liu*, X Du, X Xu, Rensselaer Polytechnic Inst., Troy, NY

TU-G-103-5 Tuesday 4:30PM - 6:00PM Room: 103

Purpose:
To develop a fast Monte Carlo code ARCHER (a testbed for Accelerated Radiation-transport Computation in Heterogeneous EnviRonments) to calculate CT imaging organ doses on heterogeneous computing systems.

Methods:
ARCHER simulates the transport of low-energy (1~140keV) photons in heterogeneous media using two selectable interaction models, simple and detailed. It includes an extensive library of voxelized human phantoms such as RPI-Pregnant women with 3, 6, 9 month gestation, extended RPI-Adult male and female phantoms representing patients of different obesity level. It contains a validated scanner model based on a GE LightSpeed third-generation 16-multi-detector CT. A series of scan protocols are predefined, including a combination of scan mode (helical or axial), beam collimation (5, 10, or 20 mm) and kVp (80, 100, 120 or 140). ARCHER is targeted for three platforms, CPU (an Intel Xeon X5650 6-core processor), GPU (Nvidia M2090 GPU), and coprocessor (Intel Xeon Phi 5110p coprocessor), and correspondingly has three code variants developed in MPI-OpenMP, CUDA and offload OpenMP respectively.

Results:
Absorbed doses to 28 organs/tissues of dosimetric interest were calculated. MCNPX was used as the accuracy benchmark. The average difference of the organ doses calculated by MCNPX and ARCHER were found to be less than 5% and 0.5% for simple and detailed interaction models respectively, indicating an excellent agreement. While ARCHER's three variants were 63~1207x faster than MCNPX, the GPU and coprocessor codes were 1.6~3.4x faster than the CPU (6 core) code.

Conclusion:
The accelerator-based new Monte Carlo code ARCHER is developed for accurate and fast CT imaging organ dose calculations. Its full capability to simulate patients of different genders and body sizes receiving different scan protocols exhibits high clinical value.

Contact Email: