Deep dose plugin: towards real-time Monte Carlo dose calculation through a deep learning-based denoising algorithm

Abstract Monte Carlo (MC) simulation is considered the gold standard method for radiotherapy dose calculation. However, achieving high precision requires a large number of histories, which time-consuming. The use computer graphics processing units (GPUs) has greatly accelerated MC and allows calculation within few minutes typical treatment plan. some clinical applications demand real-time efficiency To tackle this problem, we have developed real-time, deep learning (DL)-based denoiser that can be plugged into current GPU-based engine to enable We used two different acceleration strategies achieve goal: (1) applied voxel unshuffle shuffle operators decrease input output sizes without any information loss, (2) decoupled 3D volumetric convolution 2D axial 1D slice convolution. In addition, weakly supervised framework train network, reduces size required training dataset thus enables fast fine-tuning-based adaptation trained model radiation beams. Experimental results show proposed run in as little 39 ms, 11.6 times faster than baseline model. As result, whole pipeline finished ? 0.15 s, including both GPU DL-based denoising, needed applications, such online adaptive radiotherapy.