Room: ePoster Forums
Purpose: The dose distributions by the Eclipse electron Monte Calro (eMC) algorithm include the stochastic noises. To solve this problem, smoothing methods such as Gaussian have been used. However, these would cause distortion of the true dose distributions. Therefore, the aim of this study is to propose a deep learning method for improvement of the denoising performance in the eMC calculation process.
Methods: We used Neural Network Denoising Model with a few modifications. This model consisted of convolutional neural network (CNN) with 10 layers and U-Net. Convolutional fillers (3Ã—3), batch normalization, and rectified linear unit were applied to the CNN. Each layer of the U-net included the batch normalization and max pooling process. Residual learning was used to the output of the CNN and the U-Net. Mean square errors were used as loss. The loss was optimized by Adam algorithm. To train the model, eMC dose distributions with the stochastic noises were modeled by applying the Gaussian noise to the electron dose distributions by Hogstrom algorithm. The modeled eMC dose data were augmented to five-hundred 2D dose distribution through image rotation. Ninety percent of the data were used to training, and the remaining for test. We compared the deep learning method with Gaussian smoothing. The dose profiles and mean square errors were evaluated.
Results: Compared to the training loss, the testing loss was slightly high. In the high gradient region, the dose profiles by the deep learning were more comparable to true dose compared to the Gaussian smoothed dose profiles. In the high dose or low gradient region, the dose profiles by the two methods were similar each other.
Conclusion: The deep learning method showed better denoising performance on the dose distribution than the Gaussian smoothing. In potentially, this method could be a useful denoising tool for eMC dose calculation.
Funding Support, Disclosures, and Conflict of Interest: Funding Support: This research was supported by the Mid-career Researcher Program (2018R1A2B2005343) and Radiation Technology R&D program (2017M2A2A7A01021264) through the National Research Foundation (NRF) funded by the Ministry of Science and ICT of Korea.