Streamlining Head and Neck Auto Planning with AI-Driven Dose Prediction and Knowledge Based Plan Templates
Abstract
Purpose
Radiotherapy (RT) planning for Head and Neck Cancer (HNC) is resource-intensive and prone to variability. This study proposes and validates a fully automated pipeline synergizing deep learning-based 3D dose prediction with a knowledge-based planning (KBP) template to enhance efficiency, standardization, and plan quality.
Methods
A 3D U-Net model with attention maps was trained on 75 clinically approved HNC plans to predict 3D dose distributions from CT images and structures. Predicted doses were automatically applied to a KBP template in Eclipse, built on institutional organs at risk (OARs) constraints and refined through iterative process. Final plans were generated without manual intervention. Performance was evaluated quantitatively against clinical plans for target coverage and OARs sparing. Additionally, five planners reviewed both plan types in a blinded study using a 5-point scale: 1 (Unacceptable), 2 (Major revision required), 3 (Minor revision required), 4 (Minor stylistic change helpful), and 5 (Use as-is).
Results
The pipeline generated deliverable plans in under 30 minutes. Automated plans demonstrated superior OARs sparing, achieving a statistically significant reduction in mean parotid dose compared to clinical plans (p 3). While the remaining 15% required minor refinements in target coverage, these were easily corrected through renormalization or brief manual adjustment to optimization constraints, maintaining the overall efficiency of the workflow.
Conclusion
Our integrated pipeline automates HNC planning by combining deep learning dose prediction with a robust KBP template, reducing planning time by over 50% compared to manual planning and producing clinically usable plans with no user intervention 85% of the time. This approach delivers near-clinical quality plans while maintaining high standards of OARs sparing and workflow consistency.