ESTRO 2025 - Abstract Book

S2487

Physics - Autosegmentation

ESTRO 2025

Conclusion: We presented a comparison of different AI-models for annotation of the prostate on 2D-cine-MRI. The nnUnetV2 architecture showed the highest segmentation accuracy of the analyzed architectures. The inference speed could be decreased to 130 ms by optimizing the inference workflow and the model, allowing potentially for real-time annotation of 2D-cine MRIs at the MR-Linac in the future.

Keywords: Real-time, 2D-cine-MRI, MRgRT

2640

Digital Poster Shining a light into the Black Box: Investigating auto-contouring uncertainty Mark J Gooding 1,2 , Jonathan Lane 3 , Djamal Boukerroui 1 , Eliana Vasquez Osorio 2

1 -, Inpictura Ltd, Abingdon, United Kingdom. 2 Division of Cancer Sciences, Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom. 3 Department of Radiotherapy Physics, Oxford University Hospitals NHS Foundation Trust, Oxford, United Kingdom Purpose/Objective: There is increasing interest in assessing uncertainty of deep learning-based auto-contouring to identify regions that clinical teams should consider most carefully when editing [1]. However, most approaches, such as Monte-Carlo dropout, require access to the deep learning model. Such access is not possible for most clinics using black-box auto-contouring systems. This work aims to test the feasibility of using deformable image registration (DIR) based data perturbation to investigate auto-contouring uncertainty in a black-box scenario. Material/Methods: To perform geometric data perturbation; Each case was warped into the space of a large number of template images using a smooth, invertible registration (ITK, Insight Software Consortium), to simulate geometric variations in patient position or anatomy. These deformed images were contoured using the deep learning black-box. All auto contours were subsequently transferred back to the original case using the inverse transformations. The accuracy of the registration is of limited concern compared to the invertibility, since the requirement is to deform the case plausibly and to transfer the contours back to it, rather than to match the template accurately. The uncertainty was quantified using the 10-90% range of deviation of these contours from the auto-contour on the case itself. For this feasibility study, 124 Head and Neck CTs were used as templates in the process. An additional CT was collected from an open dataset [2], representing the case to be contoured for which the uncertainty of contouring was required. A commercial auto-contouring system (RayStation 11BR, RaySearch Laboratories, Stockholm, Sweden) was used in this study. Results: Regions with soft tissue interfaces (such as the base of the tongue) and areas where interobserver variation is known to be high [3] (such as the accessory process of the parotid) appear to have the highest level of uncertainty. Figure 1 shows the range of segmentations after the data perturbation on the original case. High uncertainty was also found at the junction between structures, such as between the spinal cord and the brainstem. Figure 2 illustrates range of contour position for the brainstem, showing uncertainty at the cranial and caudal ends.