Query-based prediction of 3D x-ray scatter fields in an interventional suite using a transformer model

Abstract Fluoroscopy-guided interventions expose clinical staff to x-ray scatter radiation, motivating accurate three-dimensional scatter-field predictions for radiation protection planning. Monte Carlo (MC) simulations provide reference accuracy but require minutes to hours of computation, while analytical models are faster but often lack spatial fidelity in complex geometries. We present a transformer-based model for scatter dosimetry in a single simulated interventional suite with fixed room geometry and object inventory, supporting near real-time inference on localised subvolumes and interactive runtimes for whole-room evaluation. We train a separate transformer-based model for each target quantity: air kerma per source photon, transported energy, and energy fluence per source photon. Each predictor encodes spectrum and technique parameters and represents room objects and their poses as a set of scene tokens. Predictions are obtained by coordinate-conditioned regression at queried voxel locations, so inference can be restricted to selected subvolumes and runtime scales with the number of queried voxels. The training set comprises 715 simulated configurations with varied C-arm poses, spectra, patient-table and patient phantom positions. On an independent test set of 48 configurations, the model achieves median R 2 > 0.996 for log-transformed targets and structural similarity around 0.83 . In linear units, median R 2 values are 0.885 for air kerma per source photon, 0.97 for transported energy, and 0.96 for energy fluence. Full-room inference, reconstructing the effective grid of about 12.2 million voxels, takes ∼ 13 s on an NVIDIA L40S GPU compared to 17 min for MC simulation, yielding a ∼ 77 × speed-up. These results represent a step toward interactive, room-scale radiation protection planning, enabling rapid what-if analyses of staff positioning and procedure planning.