Researchers have developed an innovative approach that combines Monte Carlo simulations with deep learning to improve the efficiency and precision of quality assurance in radiation therapy. This method tackles the longstanding issue of achieving both rapid computation and high accuracy in electronic portal imaging device (EPID)-based dose verification.
The Challenge in Radiation Therapy Quality Assurance
EPID serves as a vital tool for real-time in vivo dose verification during treatments. Monte Carlo (MC) simulation remains the gold standard for dose calculations, but it presents a trade-off: simulating more particles yields greater accuracy yet demands extensive computation time, while fewer particles result in noisy outputs that undermine reliability.
Integrating Monte Carlo with Deep Learning
To overcome this, experts led by Professor Fu Jin integrated a GPU-accelerated MC code called ARCHER with the SUNet neural network, designed specifically for denoising tasks. In tests using lung cancer intensity-modulated radiation therapy (IMRT) cases, the team generated EPID transmission dose data with varying particle counts: 1×106, 1×107, 1×108, and 1×109. They trained SUNet on low-particle-number data, using the high-fidelity 1×109 particle dataset as the reference standard.
Impressive Performance Gains
The combined MC-deep learning (MC-DL) framework delivered superior results in speed and dosimetric accuracy. For the noisy 1×106-particle data, SUNet denoising elevated the structural similarity index (SSIM) from 0.61 to 0.95 and boosted the gamma passing rate (GPR) from 48.47% to 89.10%. The 1×107-particle data, an ideal balance point, achieved an SSIM of 0.96 and a GPR of 94.35% post-denoising, while the 1×108-particle case reached 99.55% GPR. The denoising process took just 0.13–0.16 seconds, shortening overall computation to 1.88 seconds for 1×107 particles and 8.76 seconds for 1×108 particles. The resulting images showed reduced graininess and smooth dose profiles that preserved essential clinical details.
Advancing Clinical Applications
This technique proves especially valuable for online adaptive radiotherapy (ART), where quick dose checks help reduce patient discomfort and account for anatomical changes. It provides flexibility: 1×107 particles offer a strong speed-accuracy balance for urgent situations, and 1×108 particles ensure precision for complex cases.
“By integrating the accuracy of Monte Carlo simulation with the computational efficiency of deep learning, we have developed a practical solution that addresses the critical clinical need for rapid and reliable patient-specific quality assurance,” stated Professor Fu Jin. “This technology not only enhances existing radiation therapy workflows but also establishes a foundation for advanced applications, such as 3D dose reconstruction and broader implementation across diverse anatomical sites.”
Future efforts will extend the model to additional treatment sites, refine the SUNet architecture, and investigate other neural networks to further improve dose prediction.
