2203 - The Feasibility of EPID Transmission Image Using Monte Carlo Simulations and AI-Based Denoising

Purpose/Objective(s): This study aims to integrate Monte Carlo (MC) simulation with AI-based denoising techniques to achieve fast and accurate prediction of high-quality EPID transmission images under low-particle conditions.

Materials/Methods: A total of 100 lung cancer cases undergoing fixed-field intensity-modulated radiotherapy (IMRT) were collected. EPID transmission images were generated by the MC code under four particle history conditions (1×106, 1×107, 1×108, and 1×10?). To address the significant noise observed in low-particle simulation, the neural network model, named SUNet, was employed for denoising. The model was trained to learn the noise characteristics of low-particle images and the clarity features of high-particle images, with data augmentation techniques applied to enhance model robustness. The denoised EPID transmission images were analyzed qualitatively, and quantitative evaluations were conducted using the Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and gamma passing rate (GPR), which regards the results of the 1×10? particle as reference with 3%/2mm criteria. The computation time for both the ARCHER simulation and the denoising process was recorded.

Results: As the particle number increases, the quality of the original EPID transmission images is developed rapidly. However, the computation time required for high-particle simulation increase significantly (1×106: 1.116 s, 1×107: 1.722 s, 1×108: 8.616 s, and 1×10?: 73.888 s). In contrast, the SUNet denoising process consistently requires approximately 0.1 seconds regardless of the particle history. After applying SUNet denoising, the profiles become smoother. And the average SSIM improves from about 0.61 to 0.95 under 1×106 particles, from approximately 0.70 to 0.98 under 1×107 particles, and from around 0.90 to 0.97 under 1×108 particles. Correspondingly, PSNR increase by over 20% under both 1×106 and 1×107 conditions and by over 10% under 1×108 particles. And the average GPR ranges from about 48.47% to 89.09% under 1×106 particles, from approximately 61.04% to 94.35% under 1×107 particles, and from around 91.88% to 99.55% under 1×108 particles. These enhancements in SSIM, PSNR and GPR verify the effectiveness of the SUNet model in improving the quality of EPID transmission images obtained under low-particle conditions.

Conclusion: This study demonstrates that combining AI-based denoising techniques with MC simulation can effectively suppress noise in EPID transmission images generated under low-particle conditions while substantially reducing overall computation time. This approach provides an efficient and accurate solution for real-time verification and online adaptive radiotherapy. Future work will focus on further optimizing model performance and validating the method across a wider range of clinical scenarios.