ESTRO 2025 - Abstract Book

S2431

Physics - Autosegmentation

ESTRO 2025

1152

Digital Poster Organs-at-risk autocontouring on synthetic CTs for brain- and head & neck-MR-only radiotherapy workflows Martin Buschmann 1,2 , Harald Herrmann 1,2 , Manuela Gober 1 , Aleksandra Winkler 1 , Nicole Eder-Nesvacil 1,2 , Franziska Eckert 1 , Joachim Widder 1,2 , Dietmar Georg 1,2 , Petra Trnková 1,3 1 Department of Radiation Oncology, Comprehensive Cancer Center, Medical University of Vienna/University Hospital Vienna, Vienna, Austria. 2 Christian Doppler Laboratory for Image and Knowledge Driven Precision Radiation Oncology, Medical University of Vienna, Vienna, Austria. 3 Faculty of Nuclear Sciences and Physical Engineering, Czech Technical University in Prague, Prague, Czech Republic Purpose/Objective: MR-only radiotherapy (RT) workflows may offer increased efficiency by using a synthetic CT (sCT) generated from MR scans for treatment planning and omitting the planning CT (pCT) scan. Modern RT planning uses organs-at-risk (OAR) autocontouring, however, most commercial OAR autosegmentation tools are only available for CT, and MR based autocontouring is currently not widely commercially available. The available data-base for training MR-based autosegmentation models is also much smaller than for CT. In this study, the feasibility of OAR autosegmentation on sCT images for brain and head & neck (HN) tumors was investigated for the development of MR-only RT workflows using an established neural network-based autosegmentation software. Material/Methods: In this retrospective study, 13 brain tumor patients (glioma) and 10 HN tumor patients were included. Each patient received a pCT scan (Somatom Definition AS, Siemens Healthineers, Erlangen, Germany). An MR scan was acquired on an RT-dedicated 1.5 T MR scanner (Ingenia Ambition X, Philips, Eindhoven, The Netherlands; v5.7). sCTs were generated from a 3D T1W mDIXON sequence utilising the MRCAT (MR for Calculating Attenuation) algorithm package (Philips) which employed pre-trained convolutional neural networks. All pCTs and sCT images were sent to the CT-based autosegmentation software ART-Plan Annotate (TheraPanacea, Paris, France) for OAR autocontouring. The sCT contours were transferred to the pCT according to a 6 DoF rigid registration. Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD) were calculated. The clinical VMAT plan was recalculated on the sCT and the OAR Dmean and D1% values were compared between pCT and sCT autocontours. Results: The autocontours from brain sCTs exhibited strong geometric alignment with the pCT-based contours for large soft tissue structures, such as the brainstem, cerebellum, and entire brain, with DSC scores exceeding 0.9. Smaller brain OARs like the optic nerve, cochlea, chiasm, and pituitary showed lower DSC values, with median scores at ≤ 0.82. The HN autocontours demonstrated less agreement than brain segmentations, attributed to greater anatomical deformations between different HN scans caused by spine and mandible disalignements. The median geometric agreement was DSC=0.84 for the brain and DSC=0.79 for HN OARs. The median (±interquartile range) deviation in dose values was 0.4 (±4.1)% and 0.2 (±6.9)% for brain and HN OAR, respectively.