Fast Monte Carlo simulation of proton therapy treatment using Intel Xeon Phi coprocessor

Souris, Kevin [UCL] Lee, John Aldo [UCL] Sterpin, Edmond [UCL]

Purpose In proton therapy, range uncertainties jeopardize treatment quality. Monte Carlo (MC) simulations can help reduce them by improving dose calculation accuracy and being integrated in on-line range monitoring techniques, like prompt gamma (PG) imaging. Most current MC methods are however too slow for clinical use. Using graphical processor units (GPUs), computation times of the order of the minute have been reported. However, the highly vectorized architecture of GPUs is inadequate to handle the stochastic nature of MC methods, forcing profound adaption of the inner physics with a loss of generality. As a result, most MC codes based on GPUs are limited to a specific task, namely dose computation with the sole tracking of the protons (no PG emission). In this communication, we present MCsquare (Many-Core Monte Carlo), a new, fast, and accurate MC simulation of protons. MCsquare relies on the new Intel Xeon Phi coprocessor, which does not have the restrictions of GPUs and thus enable more general and realistic simulations. Material and methods The flexible architecture of the Xeon Phi coprocessor provides many independent execution threads that can be assigned to detailed simulations of nuclear reactions and secondary particles without performance loss. This coprocessor is composed of 60 cores, each containing a 512-bit SIMD unit, and is capable of more than 1 teraflops. In MCsquare, nuclear reactions are sampled from the ICRU 63 database, taking into account the atomic composition of tissues. ICRU 63 database provides comprehensive information for PG emission that is incorporated in MCsquare. To illustrate the accuracy and the computation speed, a 200 MeV proton beam has been simulated using MCsquare and compared to Geant4. Two cases are considered 1) a simple box of water; 2) a heterogeneous geometry integrating 3 bone material structures (see figure). The Geant4 simulation was performed with two different nuclear models, Precompound and Binary Cascade. For each experiment, 10⁶ protons were simulated. Integral depth-dose distributions and computation times were then compared. Finally, the capabilities of MCsquare to compute PG emission was compared to Geant4 (Binary Cascade) for a 150 MeV beam in an ICRP soft tissue phantom. Results MCsquare and Geant4 results are very similar (< 3% differences) for both phantoms. However, we can observe slight differences between the two nuclear models of Geant4. MCsquare results are closer to Binary Cascade at high energy and closer to Precompound at lower energy. Good agreement was also achieved for PG production (see figure). In our speed benchmark, MCsquare is up to 90 times faster than Geant4 (see table). Conclusion A new, fast, and accurate Monte Carlo code has been developed and validated with Geant4 simulations for homogeneous and heterogeneous geometries. Due to optimized transport algorithms and the use of a Xeon Phi coprocessor, the computation time is short enough for clinical use.