Author: Ruiyan Du, He Huang, Mingzhu Li, Ying Li, Hongyu Lin, Wei Liu, Shihuan Qin, Yiming Ren, Hui Xu, Lian Zhang, Xiao Zhang, Zunhao Zhang 👨🔬
Affiliation: Department of Radiation Oncology, Mayo Clinic, Medical AI Lab, The First Hospital of Hebei Medical University, Hebei Provincial Engineering Research Center for AI-Based Cancer Treatment Decision-Making, The First Hospital of Hebei Medical University, Department of Oncology, The First Hospital of Hebei Medical University 🌍
Purpose: Monte Carlo (MC) dose calculation is the gold standard in clinical CyberKnife radiation therapy (RT), considering its steep dose gradients and high-freedom non-coplanar beam angles, but extremely time-consuming. In this study, we propose a Diffusion-model-based MC calculation workflow to achieve real-time dose calculation for clinical CyberKnife RT.
Methods: A total of 117 head and neck cancer patients were randomly selected from our institute with retrospective CT and RTdose, and were further split into training (80 patients, 16160 slices), validation (20 patients, 4040 slices), and test groups (17 patients, 3,434 slices). Each patient’s treatment plan (18–40Gy prescription by 1–5 fractions) was generated using clinical Precision treatment planning system with an MC dose engine, producing low-statistic dose distributions (input), which used 1/1000 of the clinical high-statistic MC simulation particles, and high-statistic dose distributions (ground truth). Building upon baseline diffusion model (Diff-M), we compared two frameworks targeting dose distribution accuracy: (a) a Vision Transformer-based approach (DiffViT-M) focused on noise reduction and (b) a CT-driven method (DiffCT-M) leveraging physical mechanisms for enhanced anatomical precision. The output dose was evaluated via Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Root Mean Square Error (RMSE), and 3D gamma passing rates (1%/1mm, 2%/2mm, 3%/3mm) against the ground truth.
Results: For model performance, DiffViT-M showed superior accuracy, with an 11% improvement in PSNR and a 25% reduction in RMSE compared to Diff-M. It also showed excellent 3D gamma passing rates of 96.00%±0.45%, 99.22%±0.22%, and 99.98%±0.13% at the 1%/1mm/10%, 2%/2mm/10%, and 3%/3mm/10% criteria, and had an inference time of 10 ms.
Conclusion: We present a real-time, high-precision MC dose denoising framework for CyberKnife radiotherapy, leveraging a deep learning diffusion model and potentially extendable to other SBRT treatments. To our knowledge, this is the first reported application of deep learning–based MC denoising for CyberKnife.