Comparison of Two AI Auto-Contouring Solutions for Head and Neck Cancer Radiotherapy
Abstract
Purpose
Two commercial AI auto-contouring solutions, Siemens DirectORGANS (DO) and Radformation AutoContour (RAD) are compared for their structure-specific accuracy and clinical efficiency.
Methods
Twenty-two adult head-and-neck cancer patients who underwent CT simulation in 2025 were retrospectively analyzed. Sixteen structures were contoured by both AI solutions and a board-certified radiation oncologist, including mandible, brain, brainstem, oral cavity, esophagus, eyes (left/right), submandibular glands (left/right), larynx, lenses (left/right), lips, parotids (left/right), and spinal cord. Contouring accuracy was assessed using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and Mean Distance to Agreement (MDA). Statistical significance was assessed using the Wilcoxon signed-rank test (p 0.77 for both), and submandibular glands (>0.77 for both). Performance was poorest for anatomically complex or low-contrast structures, including esophagus (RAD 0.44, DO 0.45), larynx (RAD 0.56, DO 0.21), and lips (both 0.36). Comparative analysis demonstrated that RAD performed significantly better for the larynx and spinal cord across multiple geometric metrics, whereas DO show significantly better performance for the oral cavity.
Conclusion
This study demonstrates that AI-based auto-contouring performance is highly structure-dependent in head and neck radiation therapy planning. While AI performs well for well-defined structures, challenging structures continue to require manual contouring. Ongoing qualitative physician review will further inform clinical acceptability and practical implementation.