ESTRO 2025 - Abstract Book

S2524

Physics - Autosegmentation

ESTRO 2025

3377

Proffered Paper Efficient review of automatic contouring of OARs in the brain: A dual-layer quality assurance approach combining geometric and dosimetric validation Robert Poel 1 , Amith Kamath 2 , Ekin Ermis 1 , Jonas willmann 3,4 , Elias Rüfenacht 2 , Nicolaus Andratschke 3 , Peter Manser 5,1 , Daniel M Aebersold 1 , Mauricio Reyes 2 1 Department of Radiation Oncology, Bern University Hospital, Bern, Switzerland. 2 ARTORG center for biomedical research, University of Bern, Bern, Switzerland. 3 Department or Radiation Oncology, University Hospital Zürich, Zürich, Switzerland. 4 Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, USA. 5 Division of Medical Radiation Physics, Bern University Hospital, Bern, Switzerland Purpose/Objective: Deep learning (DL)-based auto-segmentation (AS) is currently clinically available for most organs at risk (OARs). AS results require expert review and adjustments according to the indication of use. Uncertainty about where errors can occur makes the review undirected, which reduces time savings. Previously explored quality assurance (QA) for AS traditional imaging features and uncertainty metrics lacked the accuracy needed for reliable QA. This study introduces a dual-layer QA approach, combining geometric accuracy with dosimetric feedback to assist the evaluation. This approach aims to improve error detection and efficiency in the evaluation process, by evaluating the dosimetric impact of segmentation variations. Material/Methods: The QA approach assesses segmentations using geometry dosimetry. Geometric validity is determined by comparing the segmentation to an independent secondary AS model, using Dice similarity coefficients (DSC) and Hausdorff distance (HD). Dosimetric impact is assessed by an in-house developed dose prediction model for glioblastoma (GBM) treatment, that predicts dose, and dose sensitivity by defining the difference between predictions on a 1 mm dilation and erosion of the concerning contour (Figure 1). Using DSC, HD, predicted dose, dose sensitivity, and dose constraints, a Swiss cheese model is used to determine if further review is required (Figure 1).