(Wednesday, 8/1/2018) 10:30 AM - 11:00 AM
Room: Exhibit Hall | Forum 4
Purpose: The accuracy of dose calculation based on small dose grid size (DGS) is higher than using large DGS. However, the dose calculation time using the small DGS takes a long time compared to large DGS. The purpose of this study is to propose a deep-learning based method that improves the DGS with less time, and to evaluate feasibility of this method.
Methods: Super-Resolution Convolution Neural Network algorithm with a few modifications was used to improve the DGS from 3.0 to 1.0 mm. The used deep-learning model consisted of the three convolution layers (9×9, 1×1, and 5×5). Rectified linear unit was applied after the convolution at each layer. Mean square error (MSE) and the Adam algorithm were used as a loss function and optimizer, respectively. This deep -learning model was trained based on 2D dose distributions. Two-hundred 2D dose distributions for two prostate cancer patients, which were calculated by VMAT plan and anisotropic analytical algorithm according to the DGSs, and 2D slices of planning-target-volume (PTV) contours were used as training data. To evaluate the feasibility, fifteen 2D dose distributions for one prostate cancer patient not used for training were used as test data.
Results: The MSE for training and test were 0.515 and 1.284 at 100 epoch, respectively. The time to convert a 2D dose distribution from 3.0 to 1.0 mm DGS was average 0.592 ± 0.009 s. In penumbra region, dose profiles for the converted dose distribution to 1.0 mm DGS were similar to dose distribution of 1.0 mm DGS than 3.0 mm. In PTV region, the converted dose distributions were underestimated than 1.0 mm DGS.
Conclusion: The proposed deep-learning method in this study improved the DGS with less time in the penumbra region. In potentially, this method could be improved through additional training data and a few modifications
Keywords
Taxonomy
TH- External beam- photons: dose computation engines- deterministic
Contact Email
