Fast Monte Carlo dose calculation based on deep learning

Monte Carlo dose calculation (MC) is the most accurate algorithm for radiotherapy, but it is very time consuming. Recently, deep learning has been introduced to speed up MC. However, reported MC networks are limited to a few cancer sites and plans of Intensity-Modulated Radiation Therapy (IMRT). This paper proposes a new network CT-MCDL to convert the low-accuracy dose (3% uncertainty) of each beam/arc to the high-accuracy dose (0.2% uncertainty). CT images were introduced as one of input channels to improve the performance covering more different cancer sites. In contrast to multiple dose maps in IMRT, Volumetric-Modulated Arc Therapy (VMAT) generates, for an arc, only one single dose map that is more complicated. The database of 43 IMRT plans and 100 VMAT plans for different sites was employed in this study. Three models of 2D MCDL, 3D MCDL and 3D CT-MCDL were trained. The 3D CT-MCDL achieved the best performance. The average 3D gamma pass rate at 2mm/2% with 2mm dose grid was 99.9% +/- 0.03% and 99.7% +/- 0.28%, respectively for IMRT and VMAT plans. Moreover, the average calculation time was accelerated about 10 times with GTX 1080. In plan optimization, the quality of the treatment plans created by using CT-MCDL for dose correction was equivalent to the quality of the treatment plans corrected by MC with 1% uncertainty. The 3D CT-MCDL proposed by this study can fast and accurately calculate dose distribution at different cancer sites and plan types (IMRT and VMAT), it can be applied clinically to dose correction in plan optimization as well as in the final dose calculation.