Physics-informed neural network with adaptive loss balancing for real-time radiotherapy dose prediction and...

Accurate and efficient dose computation lies at the heart of modern radiotherapy planning, yet practitioners are still pulled in two opposing directions: high-fidelity solvers - Monte Carlo simulation and deterministic linear Boltzmann transport equation (LBTE) solvers such as Acuros XB - deliver the physical accuracy that adaptive workflows demand, but the runtime they impose remains hard to reconcile with on-couch decision-making. The present study introduces a physics-informed neural network (PINN) framework that embeds radiation-transport-derived constraints into a 3D encoder-decoder backbone for rapid voxel-wise dose prediction and verification. The network ingests preprocessed CT volumes together with plan-specific information - gantry, collimator and couch angles, projected MLC fluence maps, and per-beam monitor units - and outputs the full 3D dose distribution. We formulate a physics residual loss inspired by the LBTE under simplifying assumptions, applied as a soft regularizer rather than as a replacement for the supervised target, so that training data calculated by the Anisotropic Analytical Algorithm (AAA, Eclipse v15.6) remains the primary supervisory signal while physical plausibility is encouraged across the volume. To overcome the gradient competition between data-driven and physics-driven terms, a momentum-based adaptive weighting strategy continuously rebalances their contributions through training. An end-to-end verification pipeline integrating gamma analysis, dose-volume histogram (DVH) evaluation, and energy-conservation checks completes prediction-plus-quality assurance in under 15 s per patient. On a held-out cohort of 27 head-and-neck and thoracic patients, the proposed system achieved mean global gamma passing rates of 97.8% at 3%/3 mm and 93.7% at 2%/2 mm (10% low-dose threshold) against the AAA reference, outperforming both a conventional 3D U-Net and a static-weight PINN; cross-checking against a subset recomputed with Acuros XB and EGSnrc-based Monte Carlo confirmed the robustness of the gain. Notably, the physics-constrained model trained on only half of the available cases matched the data-only baseline using the full training corpus - an indicator of meaningful sample efficiency. Taken together, the results suggest that combining transport-inspired regularization with adaptive training dynamics offers a practical route toward clinically deployable real-time dose verification in adaptive radiotherapy.