Application of Conditional Generative Adversarial Networks to Efficiently Generate Photon Phase Space in Medical Linear Accelerators of Different Primary Beam Parameters

Successful application of external photon beam therapy in oncology requires that the dose delivered by a medical linear accelerator and distributed within patient’s body is accurately calculated. Monte Carlo simulation currently most accurate method for this purpose but computationally too extensive routine clinical application. A very elaborate time-consuming part such generation full set (phase space) ionizing radiation components (mainly photons) to be subsequently used simulating delivery patient. We propose generating phase spaces accelerators through learning, artificial intelligence models, joint multidimensional probability density distribution properties (their location space, energy, momentum). The models are conditioned with respect parameters primary electron (unique each accelerator), which, Bremsstrahlung, generates therapeutical radiation. Two variants conditional generative adversarial networks chosen, trained, compared. also present second-best type deep learning architecture we studied: variational autoencoder. Differences between distributions obtained water phantom, test real patients using generative-adversarial-network-based Monte-Carlo-based close other, as indicated values standard quality assurance tools radiotherapy. Particle three orders magnitude faster than Carlo. proposed GAN model, together our earlier machine-learning-based tuning an MC simulator, delivers complete solution problem simulator against physical accelerator.