ESTRO 2025 - Abstract Book

S2480

Physics - Autosegmentation

ESTRO 2025

Conclusion: The nnUNet framework proves to be a robust and effective tool for automated segmentation of atelectasis and tumors in CT images, offering potential to support clinical workflows and enhance diagnostic precision. Future work will focus on expanding the dataset and incorporating multi-modal imaging for further optimization and generalization.

Keywords: Atelectasis, Lung Tumor, Segementation

2578

Digital Poster Impact of deep learning on CT-based OARs delineation for flank irradiation: a SIOP-RTSG radiotherapy panel study Mianyong Ding 1,2 , Matteo Maspero 2,3 , Semi Harrabi 4 , Emmanuel Jouglar 5 , Sabina Vennarini 6 , Timothy Spencer 7 , Britta Weber 8 , Henriette Magelssen 9 , Karen Van Beek 10 , Remus Stoica 11 , Simonetta Saldi 12 , Tom Boterberg 13 , Patrick Melchior 14 , Marry M. van den Heuvel-Eibrink 1 , Geert O. Janssens 1,2 1 -, Princess Máxima Center for Pediatric Oncology, Utrecht, Netherlands. 2 Department of Radiation Oncology, Imaging and Cancer Division, University Medical Center Utrecht, Utrecht, Netherlands. 3 Computational Imaging Group for MR Diagnostics & Therapy, Center for Image Sciences, University Medical Center Utrecht, Utrecht, Netherlands. 4 Department of Radiation Oncology radiotherapy,Heidelberg Ion Beam Therapy Center (HIT), University Hospital Heidelberg, Heidelberg, Germany. 5 Department of Radiation Oncology, Institut Curie, PSL University, Paris, France. 6 Radiotherapy Unit, IRCCS Fondazione Istituto Nazionale Tumori-Milan, Milano, Italy. 7 -, Bristol Cancer Institute, Bristol, United Kingdom. 8 Danish Center for Particle Therapy, Aarhus University Hospital, Aarhus, Denmark. 9 Department of Oncology, Oslo University Hospital, Oslo, Norway. 10 Department of Radiation Oncology, University Hospital Leuven, Leuven, Belgium. 11 -, Emergency Clinical Hospital for Children "Marie Skłodowska Curie", Bucharest, Romania. 12 Section of Radiation Oncology, Hospital Santa Maria della Misericordia, Perugia, Italy. 13 Department of Radiation Oncology, Ghent University Hospital, Ghent, Belgium. 14 Department of Radiation Oncology, University of Saarland, Homburg, Germany Purpose/Objective: Integrating deep learning (DL) in auto-contouring has markedly enhanced organ-at-risk (OAR) delineation in adult radiotherapy, yet applications in paediatrics remain limited. Existing DL models for adults often fail to translate effectively to paediatric cases [1]. This study evaluates the impact of DL-based auto-contouring with manual revision on (1) delineation time and (2) inter-observer variability (IOV) for CT-based OAR delineation in paediatric patients undergoing flank irradiation. Material/Methods: Twelve radiation oncologists from 9 countries completed a pre-meeting survey and participated in a 2-day workshop. Each participant independently delineated OARs on post-operative CTs without contrast enhancement from paediatric renal tumour patients. Standardised delineation guidelines and window/level settings were provided during the onsite training. Delineations were performed using ProKnow DS v2.0 (Elekta AB, Sweden). The OARs included heart, lungs, liver, spleen, kidneys, pancreas, and stomach-bowel. Participants were randomly divided into two groups, completing two 2-hour delineation sessions over two consecutive days. One group performed manual delineation first, followed by DL-based revision, while the other group completed the tasks in reverse order. DL contours were generated using a pre-trained nnU-Net model [2]. Five CTs (ages 1–6) were considered per group. Time (minutes) per organ delineation was recorded. Unfinished cases were accepted, but incomplete organ delineations were excluded from the analysis. Delineation accuracy was assessed using the Dice similarity coefficient (DSC) against a consensus calculated with a simultaneous truth and performance level estimation (STAPLE, threshold=0.95). IOV was quantified as the standard deviation of DSC. Key endpoints were delineation time, accuracy, and IOV.