Author: Chieh-Ya Chiu, Shen-Hao Li, Hsin-Hon Lin, Shu-Wei Wu 👨🔬
Affiliation: Department of Medical Imaging and Radiological Sciences, Chang Gung University, Proton and Radiation Therapy Center, Chang Gung Memorial Hospital Linkou 🌍
Purpose: Monte Carlo simulation enables precise calculation of dose distribution in proton therapy through tracing the radiation particles with patient tissues. However, achieving clinical-level precision requires simulating many particles, leading to long computation times. This study explores using Denoising Diffusion Probabilistic Models (DDPM) to denoise low-particle simulations, reducing computation time while maintaining high accuracy.
Methods: DDPM optimizes dose calculations by adding Gaussian noise to training images via a forward diffusion kernel (FDK) and training a reverse diffusion kernel (RDK) to remove noise and reconstruct high-precision images. The mean squared error (MSE) between true and predicted noise optimizes model parameters. The model is trained on high-particle-dose distributions generated by MCsquare, with Gaussian noise progressively added to simulate low-particle-dose scenarios. This process enables the model to learn how to predict high-particle-dose distributions from low-particle-dose inputs. The training dataset contains 9,720 liver dose distribution images simulated at various particle levels and angles, with data augmentation techniques (scaling and translation). The validation dataset includes 1,080 low-particle-dose images of liver, head and neck, and prostate cases, generated by MCsquare. Model performance was evaluated by comparing the coefficient of variation (CV) of denoised distributions with those from high-particle-count simulations, assessing its ability to enhance low-particle-dose distributions.
Results: By applying DDPM, dose distributions simulated with 270 particles achieve equivalence to those generated with 1,000 particles. This resulted in coefficient variation (CV) improvements of 29.13% for the liver, 16.15% for the head and neck, and 37.83% for the prostate, effectively reducing computation time while maintaining the CV quality of high-particle simulations.
Conclusion: The results show that DDPM effectively predicts high-particle-dose distributions from low-particle inputs, offering potential for clinical applications in achieving Monte Carlo-level precision.