"Conventional planning is manual, iterative, and can take hours or days," explains Professor Sun. "For the online AIO workflow – which aims to complete simulation, planning, and delivery in a single session – that is simply not feasible. We needed a model that could produce clinically acceptable plans within minutes, without sacrificing target coverage or organ‑at‑risk (OAR) sparing."
The team trained their model on 890 NPC cases, then refined it through four iterative versions (V1 to V4). Each version tackled a specific clinical bottleneck. V1 established a baseline using a channel‑attention 3D U‑Net for dose prediction. V2 introduced a label‑guided prioritization mechanism and hard constraints to balance tumour coverage and OAR protection. V3 improved robustness for advanced T4 tumours by incorporating a quantile loss function, enriching the training set with additional T4 cases, and adopting stochastic platform optimization. Finally, V4 focused on speed: a CT‑based Monte Carlo dose learning (CT‑MCDL) module, parallel CPU optimization, and GPU‑accelerated dose computation cut the average planning time from 15–18 minutes to just 3.5 minutes.
"Quantile loss reduces sensitivity to outliers in dose prediction, which is crucial when dealing with anatomically complex tumours," notes Professor Zhou. "And the hybrid CPU‑GPU architecture ensures that the entire optimization runs fast enough for same‑day treatment."
To test generalizability, the researchers conducted a retrospective five‑center study involving 245 patients – 125 from the development center and 120 from four external centers, each with its own imaging protocols, contouring habits, and prescription practices. AI‑generated plans were compared with expert manual plans. Despite the heterogeneity, the AI plans achieved superior or comparable dosimetric quality. Target coverage was non‑inferior or better; for most OARs, AI plans showed lower doses or non‑significant differences. At one center where manual plans prioritized OAR sparing even at the cost of target coverage, the AI model prioritized target coverage – a trade‑off that was explicitly designed and can be adjusted via the model's priority‑based customization interface.
"This tells us that the model learns robust planning strategies that transfer across different clinical environments without retuning," says Professor Jiang. "It provides a foundation for standardising plan quality across institutions."
The team then deployed the final model prospectively in 242 consecutive NPC patients treated on a CT‑linac AIO platform. Of these, 237 (97.9%) completed the online workflow; five were excluded due to technical interruptions (segmentation delays, software crashes, or network failure). Among the evaluable cases, 225 plans (94.9%) were accepted after a single automated optimization cycle – meaning the senior radiation oncologist approved the plan without any manual editing. The remaining 12 plans (5.1%) required reoptimization, mostly due to target overdose or minor contour edits.
Total planning time – from segmentation to final dose calculation – averaged 6.5 minutes (median 5.85 minutes), with the optimization step itself taking only 3.5 minutes. "This is a dramatic acceleration," emphasizes Professor Sun. "In the time it used to take just to set up a manual plan, we now complete the entire planning and QA process."
Dosimetry analysis showed excellent and consistent target coverage across all stages. Mean V100% (volume receiving 100% of the prescribed dose) was 99.4% for the primary tumour target (PGTVp), 99.3% for involved lymph nodes, and >97% for elective volumes. Even for T4 tumours – the most complex – PGTVp coverage remained above 98.9%. OAR doses met clinical constraints for critical structures (brainstem, spinal cord, optic nerves, lenses). As expected, doses to the parotid and submandibular glands were slightly elevated (mean 3,136 cGy and 5,150 cGy) due to their proximity to high‑dose targets, but still within acceptable limits.
Every AI‑generated plan underwent two independent dose verifications. Pretreatment secondary dose calculation (uAssureTx) achieved gamma passing rates >99.7% even under the strict 2%/2‑mm criterion. In‑vivo EPID‑based transit dosimetry – more sensitive to anatomical variations and setup errors – still achieved a 98.5% pass rate under the 3%/3‑mm criterion, confirming that the delivered dose matches the plan.
Crucially, the system is not a "black box". It includes a priority‑based customization mechanism that allows clinicians to manually adjust target weights before optimization. "In complex cases where a tumour abuts a critical OAR, the physician can still guide the optimization," explains Professor Zhou. "This hybrid design – standardized automation with expert override – aligns with clinical reality and builds trust."
The authors acknowledge that doses to secondary OARs such as cochleae and optic chiasm were modestly elevated in some cases, and long‑term oncologic outcomes and quality‑of‑life metrics must still be monitored. The prospective validation was single‑center; multi‑center deployment of the full end‑to‑end workflow is needed. However, the retrospective five‑center benchmark strongly supports generalizability of the planning engine itself.
"This work provides a complete development‑to‑deployment framework for AI‑driven real‑time radiotherapy," concludes Professor Sun. "It demonstrates that with careful iterative design – addressing dose trade‑offs, anatomical complexity, and computational speed – we can bring the benefits of AI directly to the patient bedside. The same methodology can be adapted to other cancer sites, accelerating the adoption of intelligent, standardized, and time‑efficient radiotherapy worldwide."
Authors of the paper include Guangyu Wang, Xin Yang, Qianxi Ni, Junxiang Tang, Kailing Huang, Hailiang Guo, Xiaobo Jiang, Wenchao Diao, Hua Li, Yuxian Yang, Lecheng Jia, Yanfei Liu, Jiaxin Deng, Kang Zhang, Danyang Li, Xiaoyan Huang, Feng Jiang, Guanqun Zhou, and Ying Sun.
This work was supported by the National Natural Science Foundation of China (grant numbers 92259202, 82441026, 12275372, and 62401636), Science and Technology Projects in Guangzhou (grant number 2024B01J1301), and the Guangzhou Municipal Health Commission (grant number 2023P-GX02). The funding agencies had no role in the design or conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; or decision to submit the manuscript for publication.
The paper "Real-Time AI-Based Radiotherapy Planning for Nasopharyngeal Carcinoma: Development and Validation" was published in the journal Cyborg and Bionic Systems on May. 18, 2026, at DOI: 10.34133/cbsystems.0544.
