ESTRO 2025 - Abstract Book

S2491

Physics - Autosegmentation

ESTRO 2025

2714

Digital Poster Assessing the long-term clinical usage of auto-segmentation for head and neck organs-at-risk Sandra van der Velden 1 , Rita Simões 1 , Mark J. Gooding 2,3 , Djamal Boukerroui 2 , Peter Remeijer 1 , Tomas Janssen 1 1 Department of Radiation Oncology, The Netherlands Cancer Institute, Amsterdam, Netherlands. 2 Inpictura Limited, _, Abingdon, United Kingdom. 3 Division of Cancer Sciences, Faculty of Biology, Medicine and Health, The University of Manchester, Manchester, United Kingdom Purpose/Objective: Auto-segmentation is often used as a starting point for organs-at-risk delineation. Manual edits to auto segmentations can vary over time due to automation bias, changes in tools, guidelines, imaging protocols or patient populations. If manual editing deviates substantially from what was observed during commissioning, centers may consider revising their tools (or clinical practice) since segmentation quality may be compromised. Continuous monitoring of auto-segmentation usage is therefore warranted [1, 2]. In this work, a commercial segmentation quality monitoring tool was employed to retrospectively assess edits made to head and neck auto-segmentations over the past 7 years, with the aim of identifying potential changes in segmentation practice that may necessitate intervention. Material/Methods: We retrospectively collected automatically generated and clinical structure sets of 1080 head and neck patients treated at our institute since 2017. Before January 2018 clinical delineations were fully manual. In January 2018 atlas-based auto-segmentation (Workflowbox, Mirada Medical Ltd., Oxford, UK) was introduced with 11 atlas patients, which was expanded to 22 atlases in July 2021. In January 2023 we clinically introduced DLCExpert, Mirada’s deep learning-based auto-segmentation. We retrospectively generated auto-segmentations for patients treated before 2018 using the first set of 11 atlas patients, to investigate the possible introduction of an automation bias. We used AIQUALIS (Inpictura Ltd., Abingdon, UK) to compare segmentations of: left and right parotid glands, left and right submandibular glands, oral cavity and glottis. AIQUALIS determined the added path length (APL) [3] and enabled the visualization of locations of the most prominent segmentation differences. Results: For all structures but the left submandibular, the APL decreased with the introduction of auto-segmentation in 2018. After expanding the atlas set, the amount of editing decreased for the glottis and the parotids (Fig.1, 2). An unexpected peak in editing was observed for the oral cavity in late 2020 (Fig.1). The introduction of deep learning led to a substantial decrease in editing for all structures but the oral cavity (Fig.1), for which the amount of editing particularly increased at the cranio-anterior side (Fig.2).