Session Item

Despite GPU acceleration, state-of-the-art Monte Carlo (MC) dose engines require considerable computation time to reduce stochastic noise. We developed a deep learning (DL) based dose engine, aiming to reach high accuracy at reduced computation times.

Following IRB approval, 350 planning CTs and RTPLANs were collected for a variety of tumor sites (step & shoot IMRT and VMAT). The GPUMCD standalone dose calculation library (Version 1, Elekta AB, Stockholm, Sweden) at a 1% (1SD) statistical uncertainty setting was used to compute 3D dose for 29100 separate segments at 6 and 10MV beam energies, both flattened and flattening filter free (FFF). A custom neural network was developed, combining  convolutions and recurrences using 50 hidden layers, taking RTPLAN parameters and 3D CT as input. Parameters were trained minimizing MSELoss between the MC computed  dose and the model output. The DL model was then used to reconstruct the full dose distribution for 6 plans not present in the training set. Gamma analyses (using local dose) were performed on the high dose region (above 50% of maximum dose) to assess accuracy. Model evaluations were performed on a PC with Nvidia RTX A4000 GPU.

Trained dose distributions in general corresponded well to the MC results, discrepancies mainly arising in case of tissue inhomogeneity (see Fig. 1). Run times depended on PTV, but were less than 30 seconds for the DL dose engine in all cases, and between 10 to 100 times shorter than the corresponding MC computation (Table 1). Gamma pass rates at 2% tolerance were above 95% for all but the IMRT lung case, in which considerable volumes of lung tissue received more than 50% of maximum dose. At 1% tolerance the gamma pass rates were considerably lower.

Fig 1. Monte Carlo (green) and DL (purple) dose distributions overlaid for a dual arc VMAT rectum case (left) and a 7 beam step & shoot IMRT lung case (right). Dose profiles along the dotted lines are shown.

Table 1. Computation times and gamma evaluation results.

A trained dose engine was implemented, delivering fair accuracy at strongly reduced computation times. Further developments will aim to reach stricter tolerances. This will support the development of real-time online correction strategies.