AI Radiotherapy Planning Cuts Treatment Time From Hours to Minutes
Researchers at Sun Yat-sen University Cancer Center developed an AI system that generates radiation treatment plans for nasopharyngeal cancer in 3.5 minutes, compared to the 15-18 minutes required by previous methods. The system completed full treatment workflows - from imaging to final dose calculation - in an average of 6.5 minutes.
The work addresses a specific clinical constraint: online adaptive radiotherapy requires simulation, planning, and treatment delivery in a single session. Manual planning cannot meet that timeline without compromising dose accuracy.
Four Iterations of Model Development
The team trained their system on 890 nasopharyngeal cancer cases, then refined it through four versions. Each version targeted a specific bottleneck.
- V1 established baseline dose prediction using a 3D U-Net architecture.
- V2 added priority-based constraints to balance tumor coverage against organ protection.
- V3 improved accuracy for advanced T4 tumors through quantile loss functions and expanded training data.
- V4 accelerated computation using CPU-parallelized optimization and GPU-based dose calculation.
The quantile loss function proved particularly important. It reduces sensitivity to outliers in dose prediction - critical when tumors sit near vital structures like the brainstem or spinal cord.
Testing Across Five Centers
Researchers compared AI-generated plans with expert manual plans across 245 patients at five institutions with different imaging protocols, contouring practices, and prescription standards. The AI system produced superior or comparable dosimetric quality without retuning.
Target coverage matched or exceeded manual plans. For most organs at risk, AI plans delivered lower doses or showed no meaningful difference. At one center where clinicians prioritized organ sparing over tumor coverage, the AI model prioritized tumor coverage - a trade-off the system allows clinicians to adjust.
Prospective Validation in 242 Patients
When deployed on a CT-linac platform for consecutive nasopharyngeal cancer patients, 237 of 242 cases (97.9%) completed the online workflow. Five cases failed due to segmentation delays, software crashes, or network issues.
Of the 237 evaluable plans, 225 (94.9%) were accepted without manual editing after a single automated optimization cycle. The remaining 12 required reoptimization, mostly due to target overdose or minor contour adjustments.
Mean tumor target coverage was 99.4% at the prescribed dose. Even for the most complex T4 tumors, coverage remained above 98.9%. Critical organ doses - brainstem, spinal cord, optic nerves - met clinical constraints.
Dose Verification Confirms Accuracy
Every AI plan underwent two independent dose verifications. Secondary dose calculation achieved 99.7% passing rates under strict 2%/2-mm criteria. In-vivo portal imaging dosimetry achieved 98.5% passing rates under 3%/3-mm criteria, confirming delivered dose matched the plan despite anatomical variations and setup errors.
System Includes Clinical Override
The system is not a black box. Clinicians can manually adjust target weights before optimization, allowing expert guidance in complex cases where tumors abut critical organs. This hybrid design - standardized automation with physician override - builds trust and aligns with clinical workflow.
The authors note that doses to secondary organs such as the cochlea and optic chiasm were modestly elevated in some cases. Long-term outcome monitoring is ongoing. The prospective validation occurred at a single center; multi-center deployment of the full end-to-end workflow remains ahead.
The methodology can be adapted to other cancer sites, the researchers said. The work provides a framework for AI-driven radiotherapy that maintains dose accuracy while meeting real-world time constraints.
Published: Cyborg and Bionic Systems, May 18, 2026.
Your membership also unlocks:
