Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
AI-Assisted Organ Delineation Improves Accuracy and Efficiency in Thoracic Radiotherapy
A prospective multicenter trial demonstrates that AI-assisted delineation of organs at risk significantly outperforms manual methods in both accuracy and speed for thoracic radiotherapy planning, while reducing inter-observer variability.
Background
Precise delineation of organs at risk (OARs) is critical in radiotherapy planning to minimize radiation exposure to healthy tissues. However, manual contouring is time-consuming and subject to inter-observer variability depending on physician expertise. This study evaluated whether AI-assisted delineation using the iCurveE deep learning model could improve clinical workflow in thoracic and breast cancer radiotherapy.
Key Findings
The prospective multicenter trial enrolled 500 patients across 5 centers with 37 physicians, generating 2,483 prospectively annotated organ sets. AI-assisted delineation significantly outperformed manual methods:
- Accuracy: Mean volumetric Dice similarity coefficient of 0.902 versus 0.857 for manual delineation (p < 0.0001)
- Distance metrics: Mean 95% Hausdorff Distance of 5.20 mm versus 8.01 mm for manual methods (p < 0.0001)
- Efficiency: 81.63% reduction in contouring time, from 55 minutes to 10 minutes median (p < 0.0001)
- Consistency: Reduced performance variability across centers and physicians of varying expertise
Why It Matters
These findings demonstrate that AI-assisted segmentation can standardize clinical practice and improve healthcare equity by providing consistent, high-quality delineation regardless of individual physician experience. The substantial time savings could enhance radiotherapy planning efficiency across institutions.
Limitations
The study’s reproducibility rating is low, and the generalizability of the iCurveE model across different imaging protocols or broader patient populations remains to be established in future clinical settings.
Original paper: A prospective multicenter trial of deep learning auto-segmentation for organs at risk in thoracic radiotherapy. — Nature communications. 10.1038/s41467-026-70863-9
