A new approach to dose estimation and in-phantom figure of merit measurement in BNCT by using artificial neural networks

Abstract

In-phantom figures of merit of the radiobiological dose distribution are the main criteria for evaluation of the boron neutron capture therapy (BNCT) plan and neutron beam evaluation. Since in BNCT there are several reactions, which contribute to the total dose of the tissue, the calculation of the dose distribution is complicated and requires lengthy and time-consuming simulations. Any changes in the beam shaping assembly (BSA) design would lead to the change of the neutron/gamma spectrum at exit of therapeutic window. As a result of any changes in the beam spectrum, the dose distribution in the tissue will be altered; therefore, another set of lengthy and time-consuming simulations to recalculate the dose distribution would have to be performed. This study proposes a method that applies artificial neural network (ANN) for quick dose prediction in order to avoid lengthy calculations. This method allows us to estimate the depth-dose distribution and in-phantom figures of merit for any energy spectrum without performing a complete Monte Carlo code (MCNP) simulation. To train the ANNs for modeling the depth-dose distribution, this study used a database containing 500 simulations of the neutron depth-dose distribution and 280 simulations of the gamma depth-dose distribution. The calculations were carried out by the MCNP for various mono-energetic neutrons, ranging from thermal up to 10 MeV energy and 280 gamma energy group, ranging from 0.01 MeV up to 20 MeV, through the SNYDER head phantom which is located at the exit of the BSA. The trained ANN was capable of establishing a map between the neutron/gamma beam energy and the dose distribution in the phantom as an input and a response, respectively. The current method is founded upon the observation that the dose which is released by the beam of composite energy spectrum can be decomposing into the various energy components which make the neutron/gamma spectrum. Therefore, in this procedure the neutron/gamma energy spectrum was converted into several energy groups and dose response of each group was predicted by the trained ANN. Total dose distribution of the entire spectrum is equal to summation of dose response of each group. If the neutron/gamma spectrum as an input changes, the dose response of that as an output can be predicted by the trained ANN in no time rather than hours or days by MCNP simulations. To check the validity of this method, this study compared full calculation of the depth-dose distribution with prediction of ANN for that. The result of this comparison shows that artificial neural networks model the dose distribution in phantom successfully and result in a great accurate prediction.