ESTRO 2025 - Abstract Book

S2495

Physics - Autosegmentation

ESTRO 2025

2839

Digital Poster Automated delineation of coronary arteries in radiotherapy planning CT using nnUnet based on data from RTOG 0617 and REQUITE trials Danai Pletzer 1,2 , Kim M. Kraus 1,3,4 , Stefan M. Fischer 1,5,2 , Lukas M. Reuter 1,2 , Julia A. Schnabel 2,5,6 , Denise Bernhardt 1,3,4 , Stephanie E. Combs 1,3,4 , Jan C. Peeken 1,3,4 1 Department of Radiation Oncology, School of Medicine at Klinikum Rechts der Isar, Technical University of Munich, Munich, Germany. 2 Institute of Machine Learning in Biomedical Imaging (IML), Helmholtz Zentrum München (HMGU) GmnH, Munich, Germany. 3 Institute of Radiation Medicine (IRM), Helmholtz Zentrum München (HMGU) GmbH, Munich, Germany. 4 Partner Site Munich and German Cancer Research Center (DKFZ), German Cancer Consortium (DKTK), Munich, Germany. 5 School of Computation, Information and Technology, Technical University of Munich (TUM), Munich, Germany. 6 Munich Center of Machine Learning (MCML), Technical University of Munich (TUM), Munich, Germany Purpose/Objective: Irradiation of the heart substructures and mainly the coronary arteries is associated with higher mortality and increased risk of adverse events in patients receiving RT for lung cancer. We sought to develop an automated model to delineate critical heart substructures in non-contrast enhanced radiotherapy planning CTs. Open-source algorithms for automated coronary artery delineation on non-contrast enhanced CT (CECT) are sparse due to their small size, limited visibility and variable anatomical course. Material/Methods: Prospective RT-planning CT data from 221 planning CT-scans from patients receiving RT for lung cancer were randomly selected from the RTOG 0617 1 and the REQUITE 3 trial. The CTs were almost equally collected from both. Manual delineation of left main (left), left circumflex (cflx) and left anterior descending (lad) coronary artery and correction was performed using 3D Slicer v5.6.0.. Models were trained using 3D nnUnets v2 5 . First, segmentations were created using platipy 2 , a hybrid model, with geometric segmentation for coronary arteries. Generated segmentations were corrected manually and served as samples for training a first nnUNet Model (model 1). The following annotations were adjusted from model´s 1 output and used for training of model 2. This was repeated for models 3 and 4 with increasing patient numbers. To remove bias towards the model segmentations, 20 samples were manually completely annotated and used for final testing. Model performance was evaluated using the DICE coefficient. For more accurate evaluation of performance on small tubular structures, the 95th percentile Hausdorff-distace (HD95%) was calculated on the final test set. Performance on the final test set was compared to two publicly available models for coronary artery segmentation, the platipy model and the nnU-Net based STOPSTORM 4 model. Results: Performance was first evaluated on validation sets and improved with the rising number of training samples from models 1 to 3. In final testing, performance improved with rising training sample size (training samples, DICE, HD95%) from model 1, to model 2, to model 3 (n=80, 0.355, 7.935mm) with best performance in model 4 (n=100, 0.358, 7.773mm) (Table 1). Compared to both the platipy and STOPSTORM model, our model performance showed a significantly higher DICE (platipy = 0.0555, STOPSTORM = 0.2754) and lower HD95% (platipy = 15.7257mm, STOPSTORM= 17.7949mm) (Table 2). Conclusion: We have developed an automated segmentation model using an iterative approach, for segmenting coronary arteries in RT planning-CT, outperforming both platipy and the STOPSTORM model, despite only being trained on non-CECT.