Development of ARCHER -- a parallel Monte Carlo radiation transport code -- for X-ray CT dose calculations Using GPU and coprocessor technologies

Abstract

Monte Carlo methods are the gold standard in radiation dose calculations with heterogeneous patient geometries and complicated irradiation conditions such as multi-detector CT scan. The long computation time has historically prevented them from becoming a routine clinic tool. The emerging hardware accelerators have created opportunities to speed up the computations significantly. They have the advantages of high computing power and high energy efficiency that are particularly suited for performing parallel tasks.

In the first test, ARCHER is in excellent agreement with MCNPX using the same geometries and similar physics. In the second test, discrepancy up to 29% from the experiment is observed, encouraging us to reinvestigate the simulation models such as the CT scanner model and the algorithm to automatically generate the phantom from CT images. In the computing efficiency test, compared to the parallel CPU code, ARCHERGPU is found to be faster by a factor of 5.40 ∼ 10.89, while ARCHERCOP is by a factor of 3.37. ARCHERGPU demonstrates good scalability when the GPU stream is implemented. The GPU platform is found to be the most energy-efficient, consuming less amount of energy than the CPU by a factor of 3.68 ∼ 8.01, while the coprocessor is better than the CPU by a factor of 2.24. Meanwhile, both the GPU and coprocessor platforms are found to be more cost effective than the CPU. Furthermore, ARCHERGPU is applied to a clinical case to compute imaging dose distributions in a patient-specific abdominal CT scan and exhibits good computing efficiency. This research shows that both the GPU and the coprocessor technology can effectively boost the performance of Monte Carlo simulations, that the GPU takes the clear lead, and that the developed code ARCHER is an important step toward patient-specific CT dose calculations.

This research represents our efforts to understand and utilize such technology in the context of radiation dosimetry, and is focused on developing and testing a new parallel Monte Carlo package, named ARCHER, for patient-specific CT dose calculations using three types of hardware platforms, including the conventional multi-core CPU and two most competitive hardware accelerators --- the Nvidia's graphics processing unit (GPU) and Intel's Xeon Phi coprocessor. ARCHER includes three variants, ARCHERCPU, ARCHERGPU and ARCHERCOP, which are tested on a 6-core Intel Xeon X5650 CPU, three Nvidia GPUs (M2090, K20, K40), and an Intel Xeon Phi 5110p coprocessor, respectively. ARCHER has a built-in model of the GE LightSpeed Pro 16 CT scanner and a library of computational human phantoms that allow realistic scan protocols to be simulated. For a fair code comparison, all the variants are carefully optimized and fine-tuned to their specific hardware platforms. Important performance factors such as the accuracy, computing efficiency, scalability and energy efficiency of the codes are investigated. The accuracy tests include the benchmark of the Monte Carlo transport kernels against the production Monte Carlo code MCNPX, and the benchmark of the simulation models against the experiment using a real human subject.