Purpose: We recently obtained nearly the same depth profiles of luminescence images of water as dose for protons by subtracting the Cerenkov light component emitted by secondary electrons of prompt gamma photons. However, estimating the distribution of Cerenkov light with this correction method is time-consuming, depending on the irradiated energy of protons by Monte Carlo simulation. Therefore, we proposed a method of estimating dose distributions from the measured luminescence images of water using a deep convolutional neural network (DCNN). Methods: In this study, we adopted the U-Net architectures as the DCNN. To prepare a large amount of image data for DCNN training, we calculated the training data pairs of two-dimensional (2D) dose distributions and luminescence images of water by Monte Carlo simulation for protons and carbon ions. After training the U-Net model for protons or carbon ions using these dose distributions and luminescence images calculated by Monte Carlo simulation, we predicted the dose distributions from the calculated and measured luminescence images of water using the trained U-Net model. Results: All of the U-Net model's predicted images were in good agreement with the MC-calculated dose distributions and showed lower values of the root mean square percentage error (RSMPE) and higher values in the structural similarity index (SSIM) in comparison with these values for calculated or measured luminescence images. Conclusion: We confirmed that the DCNN effectively predicts dose distributions in water from the measured as well as calculated luminescence images of water for particle therapy.
ASJC Scopus subject areas