French
English
Monte Carlo (MC) is considered in many Medical Physics applications as the reference simulation method: today, 100% of radiation-based systems (X-ray and nuclear imaging, linear accelerator for radiation therapy, etc) are designed by industrials and researchers with MC simulations. This is a major pillar of the field. MC is leading to the most accurate result but its convergence is generally slow. Since several years, our group investigated and proposed Variance Reduction Techniques (VRT) based on hybrid approaches mixing stochastic and deterministic approaches. Several projects supported those investigations: MC-SMART (physicancer), tGATE (ANR), SPEDIV (physicancer), Labex Primes (WP5).
A first work [Baldacci-2015- Zeitschrift fur Medizinische Physik] allowed us to introduce in GATE and characterize the Track Length Estimator (TLE) methods initially proposed by Williamson, and to propose a new, even faster, algorithm called seTLE (Split Exponential TLE) [Smekens-2014- Physics in Medicine and Biology]. This method is dedicated to the dose computation for low energy photon (lower than 1 MeV). We also studied and developed hybrid algorithms for fast simulation of Prompt-Gamma (PG) emitted during protontherapy treatment [Kanawati-2015- Physics in Medicine and Biology, Huisman-2016- Physics in Medicine and Biology]. PG are emitted by nuclear interaction during the proton interaction with matter and are currently under heavy investigation to allow on-line monitoring of delivered dose during protontherapy. Acceleration close to 103 were obtained compared to reference MC simulations. More recently, we developed a Fixed Forced Detection VRT approach for SPECT image simulation [Cajgfinger-2018- Physics in Medicine and Biology], speeding up simulation of scintigraphy by up to 2 orders of magnitude. This work makes use of software developments in both RTK and GATE [Sarrut-2014- Medical Physics]. Since a few months, our investigations on VRTs have moved towards the use of machine learning (ML) approaches during the simulation process [Sarrut-2018- Physics in Medicine and Biology]. Indeed, the probabilistic nature of MC simulations coupled with the possibility to generate a very large quantity of data led us to believe that ML could be an additional tool to improve the computation time and/or the modeling of complex physical situations (see project section).
Those advances on fast MC simulations in medical physics were also applied to small-animal RT and imaging during a project in collaborations with Nantes [Noblet-2016- Physics in Medicine and Biology]. Compared to clinical RT, pre-clinical RT systems make use of low or medium energy photon sources 100-400 kV instead of MeV, and we have shown that our methods may be well suited to decrease the computation time. It also allowed us to participate to the ESTRO[1] ACROP committee (Advisory Committee in Radiation Oncology Practice) providing guidelines for small animal radiotherapy [Verhaegen-2018- Radiotherapy and Oncology].