KEY POINTS
- This narrative review covers AI applications across proton and heavy-ion treatment planning, focusing on four major areas: automated plan generation, dose prediction and optimization, treatment adaptation, and quality assurance. It also discusses biological modeling, regulatory implementation and future computational approaches.
- Automated planning applications include selection of beam geometry, optimization parameters and fluence, with supervised learning, deep learning and reinforcement learning used to reduce iterative manual planning. The review highlights knowledge-based planning and reinforcement-learning approaches for beam-angle selection as examples.
- AI-based dose models can generate rapid three-dimensional dose estimates and potentially reduce the computational burden of conventional Monte Carlo calculation. In particle therapy, the authors emphasize that models may eventually need to predict not only physical dose but also LET and RBE distributions, where biological uncertainty remains an additional challenge.
- Adaptive particle therapy is a particularly attractive use case because proton and ion dose distributions are highly sensitive to anatomical change. AI approaches are being investigated for anticipating tumor shrinkage, organ motion and weight loss, detecting anatomical changes on daily imaging and supporting faster replanning.
- AI may also act as a quality-assurance layer by detecting abnormal plans, estimating uncertainty and supporting independent dose verification. The authors emphasize that clinically deployed systems will still require human-in-the-loop review, especially while model behavior remains difficult to explain.
- The central barriers are not model architecture but data availability, institutional generalizability, interpretability, regulation and clinical validation. Models trained at one center may fail when scanner protocols, machines, patient populations or workflows change, making multi-institutional validation and standardized benchmarks essential.
- Future directions highlighted by the review include hybrid AI–physics models, reinforcement learning, federated learning and quantum machine learning. Hybrid approaches are presented as particularly promising because they can combine the speed of learned models with physical constraints from Monte Carlo or analytical dose calculation.
CLINICAL TAKEAWAY
AI already touches almost every step of particle treatment planning, but most of the field remains at the level of technical development rather than proven clinical benefit. The near-term path is likely to be AI-assisted rather than AI-autonomous planning, with physics constraints, independent QA and multicenter validation remaining essential.