AI-Driven Radiotherapy System Plans for Nasopharyngeal Carcinoma in Minutes
Researchers have developed and validated an AI-driven real-time radiotherapy planning system for nasopharyngeal carcinoma (NPC), achieving clinically acceptable plans within minutes.
Model Development and Refinement
The model was trained on 890 NPC cases and refined through four versions (V1-V4), each addressing specific clinical bottlenecks.
V1: Baseline using channel-attention 3D U-Net for dose prediction.
V2: Introduced label-guided prioritization and hard constraints for tumor coverage and organ-at-risk (OAR) protection.
V3: Improved robustness for T4 tumors using quantile loss, additional T4 cases, and stochastic optimization.
V4: Speed optimization with CT-MCDL module, parallel CPU, and GPU acceleration, reducing planning time to 3.5 minutes.
Clinical Validation and Performance
A retrospective five-center study (245 patients) showed AI plans achieved superior or comparable dosimetric quality to manual plans.
In prospective deployment with 242 consecutive NPC patients on a CT-linac AIO platform:
- 97.9% completed online workflow
- 94.9% of evaluable plans accepted after single optimization cycle
- Total planning time averaged 6.5 minutes
Dosimetric Results
- Mean V100% >99% for primary target, >97% for elective volumes
- OAR doses met constraints
- Independent dose verification: gamma passing rates >99.7% (2%/2mm) for pretreatment, 98.5% (3%/3mm) for in-vivo EPID transit dosimetry
The system includes priority-based customization for clinician override.
Limitations
- Doses to secondary OARs (cochleae, optic chiasm) modestly elevated in some cases
- Long-term oncologic and quality-of-life outcomes pending
- Prospective validation was single-center; multi-center deployment needed
Publication
The paper was published in Cyborg and Bionic Systems on May 18, 2026 (DOI: 10.34133/cbsystems.0544).