ESTRO 2025 - Abstract Book

S2510

Physics - Autosegmentation

ESTRO 2025

homology flags insufficient contours with high (>0.85) sensitivity and excellent (>90%) specificity. This work used PG as a well-understood segmentation paradigm and may extend to target volumes and other organs-at-risk.

Keywords: Explainable artificial intelligence, parotid gland

References: 1. Strijbis, V. I. J., Gurney-Champion, O. J., Slotman, B. J. & Verbakel, W. F. A. R. Impact of annotation imperfections and auto-curation for deep learning-based organ-at-risk segmentation (IN PRESS). Phys Imaging Radiat Oncol (2024). 2. de Amorim Filho, E. C., Moreira, R. A. & Santos, F. A. N. The Euler characteristic and topological phase transitions in complex systems. Journal of Physics: Complexity (2022) doi:10.1088/2632-072X/ac664c. 3. Hacquard, O. & Lebovici, V. Euler Characteristic Tools For Topological Data Analysis. (2023) https://doi.org/10.48550/arXiv.2303.14040 4. Raudaschl, P. F. et al. Evaluation of segmentation methods on head and neck CT: Auto-segmentation challenge 2015. Med Phys (2017) doi:10.1002/mp.12197.

3152

Mini-Oral Position-Guided Auto-segmentation of Vestibular Schwannomas and Dosimetric Evaluation in Gamma Knife Radiosurgery Younghun yoon 1,2 , Shanti Marasini 2 , Alex Macintyre 2 , Eunsu Kim 2 , Chanwoong Lee 1 , Jin Sung Kim 1 , Taeho Kim 3 1 Radiation Oncology, Yonsei University, Seoul, Korea, Republic of. 2 Radiation Oncology, Washington University in St. Louis, St Louis, USA. 3 Radiation Oncology, Washington University in St. Louis, St. Louis, USA Purpose/Objective: Vestibular schwannomas (VS), the most common neoplasm of the cerebellopontine angle in adult, are recommended to obtain regular imaging for monitoring 1 . While many deep-learning auto-segmentation (DLAS) networks have been developed for VS, their performance often degraded when applied to inhomogeneous data 2 . Inspired by the anatomical origin of VS, this study introduces a positional-guided (PG) DLAS network for VS to enhance segmentation robustness and efficiency across the various imaging characteristics. Material/Methods: Three different datasets consisting of 345 VS patients who underwent contrast-enhanced T1-weighted magnetic resonance (MR) imaging were collected. Network training was conducted on 242 homogeneous public datasets 3 , while 63 institutional dataset and 40 additional public dataset were utilized as an inhomogeneous independent validation sets. The inhomogeneity included MR characteristics such as scanner, manufacturer, and sequence. Dosimetric evaluation was conducted with institutional dataset only. PG was constructed by aggregating the VS segmentations from the training set. As in Figure 1, the proposed method employed DynUnet 4 network and concatenated PGs into input channel for region-of-interest and feature extraction. The output of the network was multiplied with binarized PGs. PG was assessed by overlapping ratio with ground-truth VS from the validation set. Segmentation performance was evaluated using the dice similarity coefficient (DSC), Hausdorff distance (HD), average symmetric surface distance (ASSD), relative volume error (RVE), and false negative count (FN). Two additional networks (Unet, SwinUNETR 5 ) were employed for comparison. Dosimetric impact was analyzed using plan parameters for gamma knife stereotactic radiosurgery (GK-SRS), including coverage, selectivity, gradient index, and D95%/99%.