S1561
Physics - Autosegmentation
ESTRO 2026
brainstem and spinal cord and mean dose for swallowing-related structures. Results: Over a 33-month clinical deployment period, the AI contouring models demonstrated strong geometric and dosimetric concordance with clinician-adjusted structures. Even with stringent acceptance thresholds (SD2mm > 0.9 for critical organs, and > 0.8 for others, and mHD < 2 mm), outlier rates remained low (<2%) across most structures. The pharyngeal constrictor muscles (PCMs) showed higher outlier rates (~20%), primarily due to manual correction limitations in thin structures (as compared to image voxel size). Despite this geometric sensitivity, dosimetric differences were negligible: mean dose deviations (ΔDmean) were near zero across all structures, with 95% confidence intervals remaining narrow and symmetric (e.g., –2.3 to 2.9 Gy for the brainstem and –1.1 to 1.1 Gy for parotid glands). Even though the PCMs showed higher geometric outlier rates than other OARs, their associated dosimetric impact was minimal, underscoring the robustness and clinical reliability of the AI models across diverse anatomical regions.
Conclusion: The AI algorithm demonstrates strong geometric and dosimetric agreement with clinical data across most organs. Minimal bias and narrow confidence intervals indicate stable and reliable performance. Only select structures (the PCMs) exhibit higher variability, warranting targeted monitoring and potential model refinement. Overall, these results support continued clinical use under a structured AI performance
monitoring framework. Keywords: AI monitoring
Digital Poster 2137 The value of segmentation uncertainty for guiding expert contour correction in the prostate radiotherapy workflow Ruben J. Stoffijn 1 , Mark H. F. Savenije 2 , Cornelis A. T. van den Berg 2 , Johannes C. J. de Boer 2 , Josien Pluim 1 1 Biomedical Engineering, Eindhoven University of Technology, Eindhoven, Netherlands. 2 Radiation Oncology, University Medical Center Utrecht, Utrecht, Netherlands Purpose/Objective: Deep learning based image segmentation plays a critical role in adaptive image guided prostate cancer radiotherapy workflows. However, due to the occurrence of errors, model predictions require careful review and potential manual correction by expert therapists before being used in treatment
Made with FlippingBook - Share PDF online