ESTRO 2025 - Abstract Book

S2450

Physics - Autosegmentation

ESTRO 2025

1855

Digital Poster Fraction MRI segmentation using AI-based contour propagation in MRgRT for lung cancer Chengtao Wei 1 , Chukwuka Eze 1 , Daniela Thorwarth 2 , Cora Warda 2 , Julian Taugner 2 , Juliane Hörner-Rieber 3,4,5 , Sebastian Regnery 4,5,6 , Oliver Jäkel 7,5,8 , Fabian Weykamp 4,5,9 , Miguel Palacios 10 , Sebastian Marschner 1 , Stefanie Corradini 1 , Claus Belka 1,11,12 , Christopher Kurz 1 , Guillaume Landry 1 , Moritz Rabe 1 1 Department of Radiation Oncology, LMU University Hospital, Munich, Germany. 2 Section for Biomedical Physics, Department of Radiation Oncology, University Hospital Tübingen, Tübingen, Germany. 3 Department of Radiation Oncology, University Hospital Düsseldorf, Düsseldorf, Germany. 4 Department of Radiation Oncology, Heidelberg University Hospital, Heidelberg, Germany. 5 National Center for Radiation Oncology (NCRO), Heidelberg Institute for Radiation Oncology (HIRO), Heidelberg, Germany. 6 National Center for Tumor diseases (NCT), Heidelberg, Heidelberg, Germany. 7 Div. Medical Physics in Radiation Oncology, German Cancer Research Center, Heidelberg, Germany. 8 Department of Radiation Oncology, Heidelberg Ion-Beam Therapy Center (HIT), Heidelberg, Germany. 9 Clinical Cooperation Unit Radiation Oncology, German Cancer Research Center (DKFZ), Heidelberg, Germany. 10 Department of Radiation Oncology, Amsterdam University Medical Centers, Amsterdam, Netherlands. 11 German Cancer Consortium (DKTK), Partner Site Munich, a Partnership Between DKFZ and LMU University Hospital Munich, Munich, Germany. 12 Bavarian Cancer Research Center (BZKF), Munich, Munich, Germany Purpose/Objective: MR-integrated linear accelerators (MR-Linacs) enable high soft-tissue contrast daily imaging and online treatment adaptation. However, this can be time-consuming, with recontouring taking up to 20 minutes. Therefore, we trained a deformable image registration network for fast and accurate propagation of contours from planning to fraction MR images in MR-guided radiotherapy, preserving the personalization of organs-at-risk (OARs) and gross tumor volume (GTV) structures from initial planning. Material/Methods: A total of 140 stage 1-2 lung cancer or oligometastases to the lung patients treated at a 0.35T MR-Linac were split into 102/17/21 for training/validation/testing (labeled cohort C1 stage 1-2). Additionally, 18 central lung tumor patients treated at a 0.35T MR-Linac at an external institution (C2 stage 1-2), and 14 stage III lung cancer patients from a phase 1 clinical trial, treated at 0.35T (C1 stage 3, C3 stage 3) or 1.5T (C4 stage 3) MR-Linacs at three institutions, were used as out-of-distribution test datasets. Each patient's planning image was paired with all fraction images, and 490 pairs were used for training. Eight labels containing seven thoracic OARs and GTV were merged into one multi-label image. Hybrid transformer-CNN TransMorph models with either mean squared error (MSE), Dice, and regularization losses (TM MSE+DICE ) or with only MSE and regularization losses (TM MSE ) were trained to register planning to fraction images. The diffeomorphic models predicted dense velocity fields which were used to propagate multi-label images to generate fraction labels. Eight single-label nnUNet models were trained with planning images of the same training patients. We considered several geometric metrics and report the Dice similarity coefficient (DSC) here. Non-parametric Friedman followed by the posthoc-Nemeny test was performed between all methods.