2082 - Clinical Translation of AI-Augmented Monte Carlo Simulations: Enabling High-Resolution Proton Dose Mapping for Adaptive Therapy

Purpose/Objective(s): High-resolution Monte Carlo (MC) dose calculations are critical for precision proton therapy but remain impractical in clinical workflows due to excessive computational times. This study validates a novel AI framework that converts rapid low-resolution MC simulations into high-resolution dose maps, enabling real-time, clinic-ready dose modeling for an ultra-compact proton system.

Materials/Methods: We integrated a 3D AI model (hybrid vision transformer-convolutional architecture) into the clinical proton therapy workflow to process low-resolution MC dose maps (5mm³ voxels) generated in minutes. The model was trained on patient-specific cases (thoracic, pediatric, and head/neck tumors) and validated against a high-resolution simulation toolkit MC ground truth (1mm³ voxels). Clinical applicability was tested in heterogeneous anatomies (lung, bone interfaces, and organs-at-risk) to assess robustness in challenging scenarios.

Results: The AI framework reduced high-resolution dose calculation time from hours to <5 minutes per field, achieving 98% gamma pass rates (1%/1 mm) in homogeneous regions and 93% in heterogeneous geometries (e.g., lung tumors). Clinically significant improvements included 2.5x faster plan iteration during adaptive therapy for a pediatric medulloblastoma case. Less than 1% mean dose deviation was achieved in critical structures (brainstem, optic nerves) compared to gold-standard MC. We have also successfully modeled the penumbra to sub-millimeter accuracy, enhancing sparing of serial organs.

Conclusion: This AI-MC hybrid approach eliminates the traditional trade-off between speed and accuracy in proton therapy dose calculation. By delivering clinic-ready, high-resolution dose maps in minutes, it enables real-time adaptive replanning, robust QA for hypo-fractionated treatments, and improved organ-at-risk protection—advancing precision proton therapy for complex and pediatric cases. The framework has clinically implementation potential of seamless integration with existing workflows and actionable dosimetric improvements.