A physics-informed decomposition network for carbon ion radiotherapy dose monitoring: a proof-of-concept study

Objective. In-beam positron emission tomography (PET) provides a promising strategy for dose monitoring in carbon ion radiotherapy (CIRT), but accurate dose prediction remains difficult due to the complex, nonlinear relationship between positron-emitter activity and physical dose deposition. This proof-of-concept study aimed to improve activity-to-dose mapping by developing decomposition-based deep learning frameworks with auxiliary physical supervision. Approach. Idealized Monte Carlo (MC) simulations were conducted on computed tomography (CT) phantoms from 18 non-small cell lung cancer patients. The models were designed to predict laterally integrated one-dimensional depth-dose distributions for individual pencil-beam spots from corresponding 5 min cumulative activity and CT Hounsfield unit profiles. Two decomposition-based models, TemcoNet and NucoNet, incorporated Transformer-based decomposition modules supervised by cumulative post-irradiation activity at 10, 15, and 20 min and nuclide-specific yields of ¹¹C, ¹⁵O, and ¹⁰C, respectively, while DirectNet served as a baseline. Main results. Compared with MC ground truth, all models achieved similar median range accuracy, but TemcoNet and NucoNet substantially improved dose prediction, reducing the mean relative error from 2.36% for DirectNet to below 0.4%. The mean gamma passing rate at 2 mm/2% increased from 45.31% to approximately 96% for both decomposition-based models. Ablation experiments showed that the decomposition pathway learned physically meaningful intermediate representations, and that nuclide-yield supervision provided an additional dose-prediction benefit. Significance. Physics-informed decomposition-based modeling improves MC-derived positron-emitter activity-to-dose mapping by combining effective representation learning with auxiliary physical supervision. The proposed framework improves dose prediction while incorporating physically meaningful priors into the learning process, offering a promising basis for PET-based dose-monitoring model development in CIRT.