AI-Denoising Methods for Real-Time Multi-Particle Monte Carlo Dose Calculations Covering Photons, Protons, Carbon Ions and Neutrons
Abstract
Purpose
External-beam therapy now includes various particles such as photons, protons, carbon ions and recently neutrons in Boron Neutron Capture Therapy (BNCT). Accurate and speedy dose calculational algorithms are intergral to the clinical workflow of treatment planning and QA verification. With the advent of powerful deep-learning (DL) tools, it is possible to perform Monte Carlo (MC) dose simulations in real-time, taking a fraction of a second. This study proposes a unified real-time MC dose engine covering multiple radiation modalities using a lightweight DL-based denoising framework.
Methods
The framework integrates the ARCHER GPU-based MC dose engine with a hybrid DL denoising model. A unified lightweight architecture based on Enhanced Difference Ghost Shuffling reduces model parameters by up to 90% compared to conventional 3D-UNets while preserving high-frequency dose gradients. A strict energy-conservation constraint is incorporated to enhance physical consistency and interpretability. This model was trained and validated on a large multi-particle dataset containing over 600 3D plans for patients, including photons (190 cases), protons (175 cases), carbon ions (225 cases), and BNCT (90 cases). The Gamma Passing Rate (GPR) and Mean Absolute Percentage Error (MAPE) were used for quantitative evaluation.
Results
For photons, protons, and carbon ions, GPR (2%/2 mm) exceeded 98.7%. In BNCT cases, MAPE ranged from 2%–5% for organs at risk, with GTV dose deviations (D98, D95) below 3.5%. The method preserved sharp tumor dose gradients. Using an NVIDIA RTX 3090 GPU, the full pipeline (GPU-based simulations + AI denoising) achieved ~0.5 s for photons, protons, and carbon ions, and <27 s for BNCT neutrons.
Conclusion
By integrating a lightweight and physically constrained DL-based hybrid denoising algorithm, we have demonstrated the feasibility of extremely fast MC calculations for various types of radiation therapy particles, including neutrons in complex BNCT.