assessment of different training methods in an artificial neural network to calculate 2d dose distribution in radiotherapy

Introduction: treatment planning is the most important part of treatment. one of the important entries into treatment planning systems is the beam dose distribution data which maybe typically measured or calculated in a long time. this study aimed at shortening the time of dose calculations using artificial neural network (ann) and finding the best method of training the ann using monte carlo-n-particle (mcnp5) modeling. material and methods: back-propagation learning algorithm was applied to design the neural network. the ann was trained by mcnp5 calculations, and different kinds of methods were tested to determine the best method for training. in order to evaluate the accuracy of the ann, the beam profiles and percentage depth dose (pdd) in the field size of 15×15 cm^2 were anticipated by ann using various training methods. eventually, the results were compared with those obtained from the mcnp5 code. results: there were good agreements between the results of comparing mcnp5 calculations with experimental measurements. among the different training methods, trainbfg had the least error for calculation of pdd and beam profile. conclusion: the best training method was found to be trainbfg, and the results revealed the sufficient accuracy of the modeled ann.