Zero-Shot Cross-Modality Dose Prediction in Particle Therapy via Physics-Informed Ensembles

Abstract

Purpose: Deep learning can accelerate radiotherapy dose prediction, but particle-therapy datasets remain scarce. We evaluated whether a physics-informed ensemble trained exclusively on photon plans could predict proton and carbon-ion dose distributions without particle-dose training.Methods: Using matRad, 90 reference plans (photons, protons and carbon ions) were generated for five anatomies and six beam configurations. A 13-channel tensor combining CT, structure masks and modality-specific ray-tracing/depth-dose channels was input to SOTA-UNet3D, Res-UNet3D and Trans-UNet3D; their voxel-wise mean formed the Voltron ensemble. Leave-one-anatomy-out evaluation used four test anatomies with photon-only training. Performance was assessed with PyMedPhys gamma analysis (3%/3 mm and 2%/2 mm; 10% low-dose threshold) and mean absolute error (MAE).Results: Across four held-out anatomies, mean 3%/3 mm gamma pass rates were 23.5% for photons, 23.3% for protons and 21.5% for carbon ions; 2%/2 mm rates were 15.8%, 15.5% and 14.5%. Mean MAE was approximately 5.2–5.3 Gy. Gamma values reported in an earlier version of this work (73.2/70.8/69.1%) were computed with a non-standard lenient criterion; all values reported here use the standard gamma index.Conclusion: Photon-only training with physics-informed inputs provides a reproducible feasibility baseline for zeroshot particle-dose prediction. Performance is not clinical-grade, but the small cross-modality degradation suggests that anatomical and geometric features learned from photon data can transfer to particle therapy.