Acceleration of BNCT dose map calculations via convolutional neural networks

A carefully made treatment plan is of paramount importance in order to achieve satisfactory results in treatments based on Boron Neutron Capture Therapy. Different source configurations and positions have to be analyzed, and based on the different dose maps that can be computed, an optimal treatment should be chosen. Nowadays the dose maps are computed using slow and computationally intensive Monte Carlo simulations, which hinder the formulation of an optimized treatment plan. This work proposes a machine learning algorithm based on a convolutional neural network that accelerates the convergence of Monte Carlo neutron transport simulations, drastically reducing computation time without loss of accuracy. A dataset of Monte Carlo simulation was made and used for the training of the proposed model. 97% of the voxels of the set of testing simulations had errors lower than 5% when processed by the neural network, and inference times were reduced by three orders of magnitude. In the future, this tool could allow a real optimization of treatment plans.