Room: Track 3
Purpose: time and accurate dose engine is highly desirable in online adaptive radiotherapy. Monte Carlo (MC) based dose engine is accurate but does not meet the real-time efficiency requirement. In this work, we propose to use deep learning (DL) technique to enable real time and accurate MC dose calculation by developing a DL-based denoising plugin for the MC dose engine pipeline. This DL-based denoising model will be trained using weakly-supervised learning by using two fast and noisy MC dose distributions and compared with the results from traditional supervised learning.
Methods: dataset of 210 prostate patients with IMRT plan information were used to generate MC dose for algorithm validation. To be specific, for each patient, three MC simulations were done: two noisy simulations with1 million histories and one clean simulation with 100 million histories. The weakly-supervised learning model (Model WS) used the two noisy simulations, one as input the other as output. The traditional supervised learning model (Model TS) used one of the noisy simulations as input and the clean one as output. The 210 patients were split into 162/48, served as the training/testing datasets, respectively. A typical U-Net was used for both models. The ADAM optimizer with a 0.0003 initial learning rate was employed to train the networks for 100K iterations. The input/ouput size is of 256 by 256 by 64. The relative mean square error (rMSE) against the clean MC dose was used as the quantitative metric.
Results: noise was effectively suppressed for both models. Model WS shows comparable denoising performance as Model TS, with an rMSE of 1.9% and 1.8%, respectively. The inference speed is around 100 milliseconds for both models.
Conclusion: MC dose engine could be accelerated by 100 times by using the DL-based denoising plugin in the workflow, trained using the proposed weakly-supervised learning method.
Not Applicable / None Entered.
TH- External Beam- Photons: Computational dosimetry engines- Monte Carlo