S1599
Physics - Autosegmentation
ESTRO 2026
Germany. 3 Cluster of Excellence “Machine Learning for Science”, Eberhard Karls University Tübingen, Tübingen, Germany Purpose/Objective: Despite the growing clinical adoption of auto- contouring for organs at risk, model performance remains limited by the quality and diversity of training data, frequently failing to generalize beyond the training distribution. This issue, combined with the absence of intrinsic uncertainty estimation, increase the risk of undetected segmentation errors. Therefore, this study aimed to develop and validate an uncertainty-aware auto-segmentation pipeline that yields high segmentation accuracy on magnetic resonance imaging (MRI) of liver cancer patients and reliably identifies imperfections and missed lesions. Material/Methods: A retrospective dataset of 173 patients (n=224 liver metastases) was randomly divided into training (n=100), validation (n=33), and test (n=40) sets. Each case included T2-weighted MRI acquired on a 1.5T MR- Linac and associated clinical contours identifying liver, and single/multiple gross tumor volumes (GTV). A liver mask from a prior nnU-Net was used to extract a margin-extended hepatic region, which served as input of a SE-ResUNet for GTV delineation. To extend this framework beyond deterministic segmentation, an uncertainty-aware inference module was incorporated (cf. Fig. 1). Optimization was periodically adjusted using cyclic learning-rate scheduling to generate multiple distinct posterior weight samples. These were aggregated into entropy-based uncertainty maps, which formed the backbone of a systematic post-inference analysis – comprising detection algorithms for missed GTVs and identification methods for problematic edge areas. In both cases, preliminary contours were first generated on the uncertainty maps and subsequently projected back onto the underlying MRI for anatomical refinement. Potential findings were visualised as additional label maps (cf. Fig. 2). For model testing, ground truth contours (GTC) of 40 patients (n=52 GTVs) were compared against the AI-generated contours (AIC) using Dice Similarity Coefficient (DSC), surface DSC (sDSC; tolerance: 3mm), and 95% Hausdorff-Distance (HD95). The reliability of the uncertainty pipeline was assessed by quantifying its sensitivity in identifying annotation inaccuracies.
entire dataset, contours were predominantly edited using 2D tools such as brushes and pencils, which accounted for an average of 88.7% (±10.1) of all edits. 3D editing operations—including deformations, region growing, rotations, and translations—comprised 6.1% (±5.3) on average. Interpolations were applied in fewer than 3% of the analysed cases. For all 14,377 organs combined, the correlation between APL and editing time was 0.28. Organ-specific results are summarized in Table 1 and correlation varies from -0.15 to 0.60.
Conclusion: Audit log–based time analysis provides a reliable measure of editing time and the observed correlation between audit log–derived editing times and APL was considerably lower than that reported by Vaassen et al. (2020). Consequently, while APL remains valuable for assessing geometric differences between contours, it appears less suitable for evaluating time savings in clinical workflows. This study demonstrates that audit log data can effectively be used to estimate contour editing times, with time measured in seconds providing a more accurate reflection of editing effort than APL-based metrics expressed in millimetres. References: [1] Malone C, Nicholson J, Ryan S, Thirion P, Woods R, McBride P, et al. Real world AI-driven segmentation: Efficiency gains and workflow challenges in radiotherapy. Radiother Oncol 2025;209:110977. https://doi.org/10.1016/j.radonc.2025.110977.[2] Vaassen F, Hazelaar C, Vaniqui A, Gooding M, van der Heyden B, Canters R, et al. Evaluation of measures for assessing time-saving of automatic organ-at-risk segmentation in radiotherapy. Phys Imaging Radiat Oncol 2020;13:1–6. https://doi.org/10.1016/j.phro.2019.12.001. Keywords: contour edit time, added path length Poster Discussion 4282 Trustworthy AI: An uncertainty-aware segmentation pipeline for treatment planning of liver metastases Dominik Langner 1 , Simon Boeke 2 , Cihan Gani 2 , Daniela Thorwarth 1,3 1 Section for Biomedical Physics, Department of Radiation Oncology, University Hospital Tübingen, Tübingen, Germany. 2 Department of Radiation Oncology, University Hospital Tübingen, Tübingen,
Made with FlippingBook - Share PDF online