ESTRO 2025 - Abstract Book

S2533

Physics - Autosegmentation

ESTRO 2025

References: [1] Yves Archambault, Christopher Boylan, Drew Bullock, Tomasz Morgas, Jarkko Peltola, Emmi Ruokokoski, Angelo Genghi, Benjamin Haas, Pauli Suhonen, and Stephen Thompson. Making on-line adaptive radiotherapy possible using artificial intelligence and machine learning for efficient daily re-planning. Med Phys Int J, 8(2), 2020. [2] Kenneth Clark, et al. The cancer imaging archive (TCIA): maintaining and operating a public information repository. J. Digit. Imaging, 26(6):1045–1057, December 2013. [3] Martin Vallières, Emily Kay-Rivest, Léo Perrin, Xavier Liem, Christophe Furstoss, Nader Khaouam, Phuc Nguyen Tan, Chang-Shu Wang, and Khalil Sultanem. Data from headneck-pet-ct, 2017.

3554

Digital Poster Deep Learning for Automated Ventricle and Periventricular Space Segmentation on CT and MRI in Neuro Oncology patients Mart G.J. Wubbels 1 , Marvin Ribeiro 2 , Jelmer M. Wolternink 3 , Wouter van Elmpt 2 , Inge Compter 4 , David Hofstede 2 , Nikolina Birimac 2 , Femke Vaassen 2 , Kati Palmgren 2 , Rik H.H.G. Hansen 2 , Charlotte L. Brouwer 5 , Miranda C.A. Kramer 5 , Daniëlle B.P. Eekers 2 , Catharina M.L. Zegers 2 1 Radiation oncology (Maastro), Maastricht university medical center, Maastricht, Netherlands. 2 Radiation oncology (Maastro), Maastricht university medical center, maastricht, Netherlands. 3 department of applied mathematics, University of Twente, enschede, Netherlands. 4 Radiation oncology (Maastro), Maastricht university medical center, maastrict, Netherlands. 5 Radiation oncology, University Medical Center Groningen, Groningen, Netherlands Purpose/Objective: This study aims to create a deep learning model capable of accurately delineating the ventricles, and by extension, the periventricular space (PVS), following the 2021 EPTN Neuro-Oncology Atlas guidelines on CT and T1 weighted contrast-enhanced MRI (T1CE). The model's performance was quantitatively and qualitatively evaluated against an off-the-shelf model. Material/Methods: An nnU-Net was trained for ventricle segmentation using both CT and T1CE MRI images from 78 patients. Its performance was compared to that of a publicly available pre-trained segmentation model, SynthSeg, which was exclusively trained on a dataset of 1020 synthetic MRIs. The evaluation was conducted on both internal (N=18) and external (N=18) test sets, each consisting of CT and T1CE MRI images and expert-delineated ground truths. The nnU Net model used both CT and MRI as input, while the SynthSeg model utilized only the T1CE MRI images. Segmentation accuracy was assessed using the Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (HD95), surface DSC, and added path length (APL) using the software AIQUALIS (Inpictura Ltd, Abingdon, England). Additionally, a local evaluation of ventricle segmentations quantified differences between manual and automatic segmentations across both test sets. All segmentations were scored by radiotherapy technicians for clinical acceptability using a 4-point Likert scale. Results: The nnU-Net significantly outperformed the SynthSeg model on the internal test dataset with a DSC of 0.93 [0.86 0.95] vs 0.84 [0.69-0.89], p < 0.001; an HD95 of 0.9 [0.7-2.5] vs 2.1 [1.6-5.8] mm p<0.001; a surface DSC of 0.97 [0.90-0.98] vs 0.75 [0.63-0.84], p < 0.001; and an APL of 876 [407-1298] mm vs 3653 [2612-5667] mm, p < 0.001. Clinician ratings also favored nnU-Net segmentations on the internal set, indicating higher clinical acceptability. However, on the external test set, both models displayed no significant differences in metrics, despite a higher clinician preference score for the nnU-Net (3.6 [3.0-4.0] vs. 2.6 [2.0-3.1], p<0.001). This effect was likely due to systematic differences between internal and external test set ground truth delineations. In addition, both models showed a predisposition for high segmentation errors in the temporal horns of the lateral ventricles across the population.