Prior and Constraint-Informed Auto-Segmentation Network of Nasopharyngeal Carcinoma: A Multicenter Study
Abstract
Purpose
To prospectively validate a prior and constraint-informed deep learning auto-segmentation framework for nasopharyngeal carcinoma (NPC) across three clinical centers, addressing accuracy, efficiency, and anatomical plausibility under real-world time constraints.
Methods
A total of 176 NPC patients were enrolled from three institutions: Center A (n=126) Center B (n=20) and Center C(n=30), treated between March 2022 and November 2025. The same AI framework—comprising PCG-UNet (for GTVn), PAC-UNet (for CTV1/CTV2), and VAG-UNet (for OARs)—was deployed at all sites. All contours underwent physician review. Performance was evaluated by comparing AI contours with reference contours using Dice similarity coefficient (DSC). PCG-UNet and PAC-UNet used Gross tumor volume(GTV) as the prior and constraint information. VAG-UNet used a volume-adaptive and multi mask strategies for OARs. All models was trained by data from Center A.
Results
The framework achieved consistent high accuracy across all centers: mean DSC >0.75 for all OARs and targets (Center A: 0.911–0.862; Center B: 0.833–0.754; Center C: 0.775 - 0.762). Total segmentation time remained under 90 seconds. Anatomical constraints ensured CTV1 fully encompassed GTVp while sparing brainstem/temporal lobes, and CTV2 correctly included CTV1 without encroaching on uninvolved OARs. Volume-adaptive design improved small OAR delineation. For small OARs (volumes less than 4 cm³), the DSC was Center A: 0.852; Center B: 0.794; Center C: 0.758.
Conclusion
This multi-center prospective study demonstrates that the proposed framework delivers accurate, efficient, and robust auto-segmentation for NPC. The differences of styles between different centers result in a decrease in accuracy of models trained on single center data and evaluated in multi center. But the prior and constraint-Informed strategies still enable the models to exhibit good robustness. This segmentation framework demonstrates the potential of multi center clinical studies.