AI-Enhanced Monte Carlo Dose Verification for the World's First Fixed-Gantry Upright Proton Therapy System
Abstract
Purpose
The MEVION S250-FIT introduces a novel fixed-gantry, upright patient positioning system, yet lacks commercial secondary dose calculation (SDC) tools necessary for clinical safety verification. While Monte Carlo (MC) simulations offer gold-standard accuracy, they are computationally intensive. This study addresses these challenges by developing a validated MC model for the S250-FIT and integrating an artificial intelligence (AI) framework to generate high-resolution dose maps from rapid, low-resolution simulations.
Methods
The S250-FIT beamline—including air-core scanning magnets and adaptive apertures—was modeled in TOPAS and MCSquare. The model was validated against RayStation treatment planning system (TPS) data for upright plans (lung, prostate, brain). To accelerate calculation, we developed a U-shaped deep learning network incorporating 3D vision transformers and cross-attention mechanisms. This AI model was trained on paired low- and high-resolution MC dose maps to super-resolve noisy, low-statistic inputs into high-fidelity distributions.
Results
The foundational MC model demonstrated strong agreement with clinical measurements, with SOBP differences under 3% and average gamma analysis (3%/2mm) passing rates exceeding 90% against TPS. The AI-enhanced workflow further improved efficiency; it successfully predicted high-resolution dose distributions from low-resolution inputs, achieving over 90% gamma passing rates at the stringent 1%/1mm criteria.
Conclusion
This work presents the first AI-integrated MC dose verification system for upright proton therapy. By combining robust beamline modeling with deep learning-based super-resolution, we provide a clinically feasible, high-accuracy solution for patient safety verification.