Integrated Monte Carlo and deep learning improves radiotherapy QA

Bridging speed and accuracy in radiotherapy QA

The research, led by Professor Fu Jin, addresses a key challenge in radiotherapy: balancing computational speed and accuracy for EPID-based dose verification. EPID has emerged as an important tool for real-time in vivo dose verification. However, MC simulation, long considered the “gold standard” for dose calculations, faces a dilemma. Increasing the number of simulated particles guarantees higher accuracy, but significantly increases computation time. On the other hand, reducing the number of particles introduces destructive noise that makes the results unreliable.

Integrated MC-DL technology

To address this challenge, the team combined the GPU-accelerated MC code ARCHER with the SUNet neural network, a sophisticated deep learning architecture focused on denoising. Using a lung cancer IMRT case, they first generated noisy EPID transmission dose data with four different particle counts (1×10⁶, 1×10⁷, 1×10⁸, and 1×10⁹) via ARCHER. SUNet was then trained to denoise low particle count data using a high-fidelity 1×10⁹ particle dataset, which served as the gold standard reference for monitoring.

Remarkable results: Achieving speed and accuracy

The integrated MC-DL framework demonstrated superior performance in both computational speed and dosimetric accuracy. When processing the originally noisy 1×10⁶‑particle data, SUNet denoising improved the structural similarity index (SSIM) from 0.61 to 0.95 and increased the gamma passage rate (GPR) from 48.47% to 89.10%. For the 1×10⁷‑particle dataset representing the optimal trade-off, the denoising results achieved SSIM 0.96 and GPR 94.35%, while for 1×10⁸‑particles it reached GPR 99.55% after processing. The denoising step itself required only 0.13–0.16 seconds, reducing the total computation time to 1.88 seconds for the 1×10⁷‑particle level and 8.76 seconds for the 1×10⁸‑particle level. The denoised images exhibited smooth dose profiles with significantly reduced graininess and preserved clinically relevant features, confirming the practical feasibility of this approach for efficient QA in radiotherapy.

Enhance clinical practice and future research

This advancement has particular implications for online ART, where rapid dose verification is essential to minimize patient discomfort and reduce anatomical changes during treatment. This method provides a flexible solution. 1×10⁷ particles provide the best balance between speed and accuracy for time-sensitive scenarios, while 1×10⁸ particles provide higher accuracy for demanding 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 of rapid and reliable patient-specific quality assurance,” said Professor Fu Jin. “This technology not only enhances existing radiotherapy workflows, but also establishes the foundation for advanced applications such as 3D dose reconstruction and widespread implementation across diverse anatomical sites.”

The team plans to extend this model to other treatment sites, further optimize the SUNet architecture, and consider additional neural network approaches to improve dose prediction capabilities.