ESTRO 2025 - Abstract Book

S277

Brachytherapy - Gynaecology

ESTRO 2025

Conclusion: AI-assisted contouring demonstrates potential for improving treatment planning in brachytherapy by achieving similar or better dosimetric results compared to manual contouring, especially in terms of OAR sparing. While the differences in OAR dosimetry are generally small, the AI approach offers benefits in consistency and efficiency. A hybrid approach that integrates clinician oversight remains essential to optimize treatment accuracy and ensure the highest clinical standards. Overall, AI-driven tools could enhance the precision and workflow of brachytherapy planning while maintaining high clinical standards. References: 1.Tao Liu, Shijing Wen, Siqi Wang, Qiang Yang, Xianliang Wang, Artificial intelligence in brachytherapy, Journal of Radiation Research and Applied Sciences, Volume 17, Issue 2, 2024, 100925, ISSN 1687-8507. 2.Chen, J. et al. (2024) A review of Artificial Intelligence in Brachytherapy, arXiv.org. 3. Susovan Banerjee, MD, Shikha Goyal, MD, DNB, Saumyaranjan Mishra, MD, Deepak Gupta, MD, Shyam Singh Bisht, MD, Venketesan K, MSc, Kushal Narang, MD, Tejinder Kataria, MD, DNB. Artificial intelligence in brachytherapy: a summary of recent developments British Journal of Radiology , Volume 94, Issue 1122, 1 June 2021, 20200842. Proffered Paper Design and internal evaluation of a Deep Learning-based automatic CT segmentation of gynaecological brachytherapy Norina Predescu 1 , Monica E Chirila 1,2 , Szabolcs B Lőrincz-Molnár 1 , Saad U Akram 1 , Gregory Bolard 1 , Jarkko Niemelä 1,3 1 Department of Clinical Development, MVision AI, Helsinki, Finland. 2 Department of Radiation Oncology, Amethyst Radiotherapy Centre, Cluj-Napoca, Romania. 3 Department of Medical Physics, Turku University, Turku, Finland Purpose/Objective: Recent studies reported time savings and increased consistency when deep-learning-based auto-contouring solutions were used for pelvic organ-at-risk delineation in external beam radiotherapy planning (1-4). Reducing contouring time and interobserver variation in gynecologic brachytherapy (BT) remains a clinical challenge. Deep learning-based auto-contouring could decrease patient waiting time with the applicator in place. Material/Methods: A total of 406 female pelvis CT scans were collected to develop the auto-segmentation model, comprising 346 scans from 7 European clinics and 60 from public datasets. A team of medically trained professionals annotated the anatomical structures following international guidelines: RTOG 2012 for bladder and utero-cervix, ESTRO ACROP 2018 for rectum, GHG 2020 for sigmoid, bowel loops and urethra, and AIRO 2023 for vagina. Brachytherapy applicator delineation was also included. Quality assurance included peer review by at least two team members for training scans, with validation scans receiving additional review from an external experienced clinician. The scans were utilized to train a deep learning model based on a 3D encoder-decoder U-Net architecture. Twenty test scans, selected to ensure anatomical and geographical diversity, were held out from the training set. Model performance was evaluated by comparing automated versus manual contours using Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and surface DSC with 2mm tolerance (sDSC2). Results: The model achieved high accuracy for critical organs at risk with median DSC values of 0.96 (bladder), 0.89 (rectum), 0.86 (bowel loops) and 0.80 (sigmoid) (Figure 1). Corresponding median HD95 values were 3mm, 6mm, 9mm, and Keywords: Brachytherapy, auto-contouring, dosimetry. 4483