A Hybrid Monte Carlo and Deep Learning Approach for Fast and Generic Photon Beam Dose Calculation

Dosimetry is an essential tool to provide the best and safest radio-therapies to a patient. In this field, Monte-Carlo simulations are considered to be the golden standard for predicting accurately the deposited dose in the body. Such methods are very time-consuming and simpler dose calculation engines, like dose kernel approaches, were created for cases where a fast estimation is necessary. In those approaches, dose distribution maps (or dose kernel) are learned for simple beams geometry, then, a combination of numerous kernels is used to simulate more complex beams. However, those methods often lack personalization and oversimplify the human body, especially for the secondary interactions dose deposition. In this article, we explore the possibility of learning the dose kernel using convolutional neural networks to improve their accuracy towards each different human body. We also highlight the limits of such approaches.