S1569
Physics - Autosegmentation
ESTRO 2026
(range 0.81 – 0.89) and 0.83 (range 0.75 – 0.86), respectively. Across all nodal regions, MPC-derived delineation error was significantly smaller than that observed in the IOV study (p < 0.0001), with median differences of 1mm. Directional patterns ofuncertainty were strongly preserved. Spearman correlations between MPC- and IOV-derived delineation error were high: 0.80 (para-aortic), 0.91 (common iliac), 0.65 (pre- sacral), 0.92 (external/internal iliac) and 0.98 (obturator), indicating that MPCs captured the anisotropic nature of observer variability. Most of MPCs lay within the IOV-derived range, whereas fewer fell within the narrower clinician-defined range (Table 1).
Conclusion: Training a stochastic segmentation network on clinician-defined uncertainty ranges generated MPCs that reproduced the directional patterns of delineation uncertainty, with magnitudes falling between the conservative clinician-defined ranges and the overly broad IOV. This offers a promising approach for producing clinically meaningful delineation Baumgartner CF, Tezcan KC, Chaitanya K, Hötker AM, Muehlematter UJ, Schawkat K, et al. PHiSeg: Capturing Uncertainty in Medical Image Segmentation. Medical Image Computing and Computer Assisted Intervention – MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II; Shenzhen, China: Springer-Verlag; 2019. p. 119–27.2. Tudor G, Bernstein D, Riley S, Rimmer Y, Thomas S, Herk M, et al. Geometric Uncertainties in Daily Online IGRT: Refining the CTV-PTV Margin for Contemporary Radiotherapy. 2020. Keywords: Uncertainty quantification, Delineation error Poster Discussion 2279 Predicting the dosimetric impact of local manual adjustments to organ of interest auto- segmentations in HNC uncertainty. References: 1. Joëlle E. van Aalst 1 , Tomas M. Janssen 2 , Federica C. Maruccio 2 , Rita Simões 3 , Johannes A. Langendijk 1 , Peter M.A. van Ooijen 1 , Charlotte L. Brouwer 1 1 Radiation Oncology, University Medical Center Groningen, Groningen, Netherlands. 2 Radiation Oncology, Netherlands Cancer Institute / Antoni van Leeuwenhoek, Amsterdam, Netherlands. 3 Radiation Oncology, Netherlands Cancer Institute / Antoni van Leeuwenhoek, Groningen, Netherlands Purpose/Objective: Clinically used auto-segmentation models need manual oversight. Manual adjustments are typically guided by resource and time-intensive expert judgment. However, often edits are made with limited dosimetric impact. This is because it remains challenging for experts to judge the impact of an adjustment, which depends on both the edit itself and the clinical context. Previous studies have identified predictive features of dosimetric effects for global segmentation errors, such as organ of interest (OOI) distance to PTV [1] and local dose [2,3]. This study aimed to develop a classification model to predict if local segmentation edits lead to a change in OOI mean dose (Dmean) by combining previously identified
Made with FlippingBook - Share PDF online