Multimodal AI proposed major surveillance reduction after stage II NPC
External validation achieved AUC 0.986, while modeled risk-adapted surveillance reduced visits by 90.0–98.5% and identified every observed failure.
Tag
AI
AI is moving across particle therapy planning but clinical validation remains the bottleneck
AI can automate particle planning, dose prediction, adaptation and quality assurance, but limited data and validation still constrain clinical adoption.
Deep learning predicted distant metastasis from CT and dose in head and neck cancer
A model integrating planning CT, 3D dose and tumor contours achieved strong external prediction of distant metastasis but weaker local-recurrence performance.
AI biomarker substantially reduced planned ADT use in intermediate-risk prostate cancer
ASTuTE biomarker testing changed 27.5% of shared decisions and reduced planned short-term ADT use from 37% to 12.5%.
Generative AI is spreading across radiation therapy workflows, but validation remains the bottleneck
A scoping review maps generative AI across imaging, documentation, QA and decision support while emphasizing major safety and governance gaps.
Commercial AI synthetic CT showed good soft-tissue agreement but important high-density failures
AI synthetic CT preserved soft-tissue HU and geometry well, while bone, air interfaces and implanted devices remained important failure modes.
An LLM agent generated head and neck IMRT plans comparable to clinical plans
Reference-free LLM planning improved hotspot control and boost-target conformity while maintaining organ-at-risk sparing comparable to clinical plans.
Fifteen years of radiotherapy data were converted into an AI-ready resource
An open-source workflow structured 15 years of radiotherapy data, although reirradiation prediction showed only modest discrimination.
Transformer dose engine achieved near-Monte Carlo accuracy for carbon ion therapy
The Transformer calculated carbon-ion pencil-beam doses in 14 milliseconds with 98.0% gamma agreement at stringent 1%/1 mm criteria.
AI-assisted VMAT planning shortened workflow but required manual hotspot correction
AI-assisted planning cut average planning time from about two hours to 30 minutes, but three of 12 plans required manual hotspot correction.
Reinforcement learning adapted proton spot delivery across respiratory phases
Patient-specific reinforcement learning improved moving-target coverage over static GTV plans while reducing some normal-tissue exposure compared with ITV planning.
GPT-4.1 retrieved TrueBeam troubleshooting history but retained safety risks
A retrieval-augmented GPT-4.1 chatbot rapidly surfaced previous TrueBeam faults, but procedural, role-definition and safety errors remained.
High geometric agreement did not guarantee physician acceptance of AI contours
Thoracic AI contours showed strong geometric and dosimetric agreement, but physicians still rejected some heart and oesophageal contours.
Federated learning improved radiation pneumonitis prediction across lung cancer datasets
A center-specific federated learning model predicted grade 2 or higher radiation pneumonitis with stable cross-dataset performance.
Artificial intelligence planning generated clinically acceptable whole-brain radiotherapy plans
Artificial intelligence planning produced clinically acceptable whole-brain radiotherapy plans in 14 of 15 independent test cases.