Abstract ID: 133 Fast and accurate 3D dose distribution computations using artificial neural networks

In radiation therapy, the trade-off between accuracy and speed is the key of the algorithms used in Treatment Planning Systems (TPS). For photon beams, commercial solutions generally relies on analytic algorithms, biased Monte Carlo, or heavily parallelized Monte Carlo on Graphics Processing Units (GPU). Alternatively, we propose an algorithm using Artificial Neural Network (ANN) to compute the dose distributions resulting from ionizing radiations inside a phantom [1] , [2] . We present an evolution of this platform taking into account modulated field sizes and shapes, and various orientations of the beam to the phantom. Firstly, tomodensitometry-based phantoms are created to validate the dose distribution computed for a square beam in heterogeneous areas (head and neck, lungs). Secondly, IMRT treatments are simulated in these phantoms. To validate our approach, we compare our results with the Analytical Anisotropic Algorithm (AAA) and Monte Carlo simulations. Cross-comparisons are performed for square beams and IMRT treatments. The dose distributions are evaluated using gamma indices and profile extractions. The dose distributions computed from IMRT treatments require less than two minutes using a standard Central Processing Unit (CPU). We aim at providing a fast and accurate solution for TPS quality control.