Monte Carlo–Level Dose Prediction for Clinical Cyberknife Radiotherapy Using a Physics- and Spatially-Informed Diffusion Model
Abstract
Purpose
Monte Carlo (MC) dose calculation is the gold standard for CyberKnife radiotherapy, but its clinical integration is hindered by prohibitive computational latency. We proposed a physics- and spatially-informed diffusion model with Vision Transformer (PSIDMV) to achieve rapid, MC-level dose prediction.
Methods
Data from 251 cancer patients (head-and-neck, lung, and liver) were divided into training, validation, and testing cohorts (8:1:1). The PSIDMV model integrates Finite Size Pencil Beam (FSPB) doses as physical priors, CT images, and multi-channel signed distance maps (SDMs) as spatial priors. A Vision Transformer (ViT) was integrated into the diffusion backbone to model long-range spatial dependencies relevant to radiation transport related dose deposition. Model performance was evaluated using Mean Absolute Error (MAE) and 3D Gamma passing rates (GPR).
Results
The PSIDMV model significantly outperformed FSPB and baseline models (P < 0.05). Across all anatomical sites, MAE was reduced by more than 85% compared with FSPB. For the 1%/1mm/10% criterion, GPR reached 98.00%±2.50% (head-and-neck), 94.00%±1.50% (lung), and 95.00%±1.80% (liver). The model achieved a 3.5-fold speedup over clinically configured MC simulations, reducing computation time to 18 minutes.
Conclusion
The PSIDMV model effectively synergizes physical and spatial priors, providing a robust and efficient tool for high-precision dose calculation and secondary verification in CyberKnife radiotherapy.