S1568
Physics - Autosegmentation
ESTRO 2026
Sooha Kim 1,2 , Omar Todd 3 , Raghav Mehta 3 , Fabio De Sousa Ribeiro 3 , Katherine Mackay 4 , David Bernstein 5,6 , Ben Glocker 3 , Alexandra Taylor 2,1 1 Radiotherapy and imaging, The Institute of Cancer Research, London, United Kingdom. 2 Gynaecology Unit, The Royal Marsden NHS Foundation Trust, London, United Kingdom. 3 Department of Computing, Imperial College, London, United Kingdom. 4 Clinical Oncology, Imperial College Healthcare NHS Trust, London, United Kingdom. 5 Joint department of Physics, The Royal Marsden NHS Foundation Trust, London, United Kingdom. 6 Joint department of Physics, The Institute of Cancer Research, London, United Kingdom Purpose/Objective: Artificial intelligence (AI)–based auto-contouring has improved radiotherapy efficiency and consistency, but most models generate a single ‘best-fit’ contour. Existing uncertainty quantification methods often fail to reproduce delineation uncertainty, a major contributor to systematic geometric error. This study introduces and evaluates a novel approach of training a stochastic segmentation network on clinician- defined range contours (rather than a single contour per case) capable of generating multiple plausible contours (MPCs) to better capture clinically meaningful delineation uncertainty (1). Material/Methods: A bespoke dataset of 55 cervical cancer cases was created, each containing a three-tiered hierarchical structure set: an ‘inner contour’ containing voxels with high confidence of belonging to the structure, a standard clinical contour and an ‘outer contour’ representing the outermost plausible boundary. The elective nodal region was selected as the test structure. A stochastic segmentation network was trained to produce MPCs. Performance was evaluated against a STAPLE consensus contour derived from an inter-observer variation (IOV) study conducted on six evaluation cases. Agreement with clinician-defined inner and outer contours was also measured. Direction-specific delineation error was estimated using a range-based approximation across eight axial directions for five nodal regions, and compared between MPCs and IOV using the Wilcoxon signed- rank test (2). Directional uncertainty patterns were examined using Spearman correlations. The proportion of MPCs falling within IOV- and clinician- defined ranges was calculated. Results: The model was set to produce 16 plausible contours and a mean predictive contour per case. Agreement with the STAPLE contour was high, with Dice values of 0.85–0.88 across six evaluation cases. Median Dice similarity between mean prediction contours and clinician-defined inner and outer contours was 0.85
Conclusion: The results demonstrate that effective CBCT-based auto-segmentation models can be trained without large-scale segmented CBCT datasets by leveraging image synthesis models. The reduced performance of CBCT- and sCT-based models is likely due to poorer CBCT image quality and reduced soft-tissue contrast. The proposed workflow enables development and training of large-scale multi-organ CBCT auto- segmentation models for adaptive radiotherapy. References: References[1] Vestergaard, C. D., Elstrøm, U. V., Muren, L. P., Ren, J., Nørrevang, O., Jensen, K., & Taasti, V. T. (2024). Proton dose calculation on cone-beam computed tomography using unsupervised 3D deep learning networks. Physics and Imaging in Radiation Oncology, 32, 100658.[2] Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2), 203-211.[3] This work was funded by a grant from Deutsche Krebshilfe (70114849) Keywords: CBCT autosegmentation, image synthesis Digital Poster Highlight 2267 AI-Generated Multiple Plausible Contours for Clinically Meaningful Uncertainty Quantification in Auto-Contouring
Made with FlippingBook - Share PDF online