ESTRO 2025 - Abstract Book

S2474

Physics - Autosegmentation

ESTRO 2025

[5] F. Isensee et al. , nnU-Net Revisited: A Call for Rigorous Validation in 3D Medical Image Segmentation, https://arxiv.org/abs/2404.09556v2

2484

Digital Poster Auto-segmentation for reirradiation: comparison with the de-novo setting in head and neck cancer Chelmis M Thiong'o 1 , Marcel van Herk 1,2 , Matthew Lowe 1,3 , Ane Appelt 4 , David Thomson 5 , Eliana Vasquez Osorio 1,2 1 Division of Cancer Sciences, The University of Manchester, Manchester, United Kingdom. 2 Radiotherapy Related Research, The Christie NHS Foundation Trust, Manchester, United Kingdom. 3 Christie Medical Physics and Engineering, The Christie NHS Foundation Trust, Manchester, United Kingdom. 4 Leeds Institute of Medical Research at St James's, University of Leeds, Leeds, United Kingdom. 5 Clinical Oncology, The Christie NHS Foundation Trust, Manchester, United Kingdom Purpose/Objective: Estimating cumulative dose metrics in head and neck reirradiation requires highly consistent contours. Auto segmentation tools have the potential to enable standardised and consistent contouring, expand the number of organs at risk (OAR) considered for dose evaluation, and save time. However, the accuracy of these tools often depends on the representativeness of the data used during model training [1] . In the reirradiation setting, large anatomical changes are common, often caused by previous radiation courses and/or other interventions. Consequently, auto-segmentation tools need to be evaluated in this setting. Material/Methods: For 27 patients with head and neck cancer who received reirradiation, OAR contours were generated using two commercial deep-learning-based auto-segmentation tools (Limbus Contour v 1.7.0-B3, Limbus.AI, Saskatchewan, Canada and RSL Head and Neck CT model within RayStation 11B-R, Release 12.0, RaySearch Lab, Stockholm, Sweden, arbitrarily referred to hereafter as Tool1 and Tool2. We assessed geometrical similarity using mean distance to agreement (mDTA), and their dosimetric impact by comparing dose estimates (mean/max) derived from manual/autosegmented OARs. Only OARs where both the clinical and auto-segmented contours were present on the planning CTs of the de-novo and reirradiation courses were considered for dosimetric assessment. We quantified the consistency of the dosimetric estimates using Bland Altman analysis for those contours that met an acceptance threshold of mDTA≤2mm. Results: Overall, 50.4% and 46.0% of OARs met the mDTA threshold for Tool1 and Tool2, respectively. This modest performance is likely due to inconsistencies in manual contours, Fig.1. The reirradiation auto-segmented OARs had larger mDTA than the de-novo OARs in 51.8% of cases, indicating lower similarity to manual contours, Fig. 2a. For structures with mDTA≤2mm, mean/max dose differences between manual and auto-segmented OARs were slightly larger in the de-novo setting, as shown by the generally wider bands in the Bland-Altman graphs, compared to the reirradiation setting, Fig. 2b. Tool1 had narrower bands.