Real-Time AI-Based Radiotherapy Planning for Nasopharyngeal Carcinoma: Development and Validation

Background: Online all-in-one (AIO) radiotherapy workflows enable same-day treatment by integrating simulation, planning, and delivery into a single session. However, for anatomically complex tumors such as nasopharyngeal carcinoma (NPC), generating high-quality plans within strict time constraints remains a major barrier to clinical adoption. Methods: We developed a deep-learning-based automated planning model specifically tailored for real-time NPC planning in the online AIO workflow. The model was trained on 890 patients and iteratively refined through 4 versions, incorporating innovations such as quantile loss, priority-based constraint encoding, and hybrid central processing unit-graphics processing unit acceleration. Model performance was benchmarked in a 5-center retrospective study, including 125 patients from the model development center and 120 patients from 4 external centers. It was then prospectively validated in 242 consecutively treated patients with NPC using a CT-linear accelerator-based AIO platform. Results: In the 5-center retrospective evaluation, artificial intelligence (AI)-generated plans achieved superior or comparable dosimetric quality relative to expert manual plans, despite variations in imaging, contouring, and prescription practices. In prospective deployment, 95% of plans were clinically accepted after a single optimization cycle, with a mean generation time of 3.5 min. All plans met target coverage criteria and passed both secondary dose verification and in vivo electronic portal imaging device analysis. Conclusion: This study represents the largest prospective validation to date of AI-based treatment planning for NPC, demonstrating real-time feasibility, robust generalizability, and consistent clinical quality. Our development-to-deployment framework supports the scalable adoption of AI-driven precision planning and provides a transferable model for intelligent radiotherapy across disease sites.