Gpu Accelerated Monte Carlo Simulation For Radiotherapy Dose Calculation

The Monte Carlo (MC) method has been the most accurate for radiation dose calculation so far. However, the large amounts of computing time necessary for employing MC technique have limited its use in clinical practice. Graphic Processor Units (GPUs) have recently demonstrated substantial parallel computing power to accelerate scientific computing applications. The purpose of this study is to speed up MC simulation for photon/electron radiation dose calculation by using GPU on a normal desktop PC, and to test the feasibility of replacing the Cluster of Workstations (COW) that is high cost, not widely available, and not easy to use for MC dose calculation. We developed our program that is functionally equivalent to standard EGS4 by using a new data-parallel, C based CUDA programming language API for NVIDIA graphic boards. The hardware employed was a NVIDIA GeForce 9800 GX2 installed on a normal desktop PC. The NVIDIA 9800 GX2 has 32 Multiprocessors with 8 processors each adding up to 256 processors in total, and 1 GB of global memory. We use 256 blocks and each of them has 256 threads so in total one grid has 65536 threads. Each thread processes one particle history for photon/electron energy reduction and dose deposition, so there are 65536 particle histories being parallel calculated with one time kernel function call. Kernel function is executed in a serial way until all particle histories are processed. We tested our program on a simple two-layer water phantom (20 cm x 20cm x 20cm) by simulating 20MeV, 15MeV, and 6MeV parallel electron beam, each involving 10e5, 10e6, and 10e7 histories (respectively). The depth dose curves produced from our program showed a very good agreement with those from the EGSnrc. But for efficiency, our program significantly improved the performance and outperformed than the EGSnrc on the COW of six personal computers in our lab. This study shows that MC simulation for dose calculation is very suitable to be implemented by using CUDA interface for NVIDIA GPUs on a normal desktop PC. The performance can be significantly speed up by this new technique. The simulation of clinical linear accelerator on patient data needs to be further investigated.