ESTRO 2025 - Abstract Book

S247

Brachytherapy - Gynaecology

ESTRO 2025

Conclusion: Although a retroverted uterus during ICBT was associated with lower CTV doses compared to an anteverted uterus, the local control and survival rates were similar. However, the presence of a retroverted uterus caused an increase in rectum and bowel doses, leading to a higher rate of grade 1-2 late GI toxicity, yet no difference in severe toxicity rates.

Keywords: anteverted, retroverted

2561

Proffered Paper Deep learning for fast and accurate automated reconstruction of multiple applicators used in MRI-based cervical cancer brachytherapy Beatriz M Fernandes 1,2 , Ioannis Androulakis 1 , András Zolnay 1 , Inger-Karine Kolkman-Deurloo 1 , Brígida Ferreira 3 , Theo van Walsum 4 , Ben Heijmen 1 , Linda Rossi 1 1 Department of Radiotherapy, Erasmus MC, Rotterdam, Netherlands. 2 Department of Physics, Faculty of Sciences of the University of Lisbon, Lisbon, Portugal. 3 Institute of Biomedical Engineering and Biophysics, University of Lisbon, Lisbon, Portugal. 4 Department of Radiology & Nuclear Medicine, Erasmus MC, Rotterdam, Netherlands Purpose/Objective: Brachytherapy for advanced cervical cancer is a time-intensive process with patients suffering from anxiety, pain and discomfort [1]. Recently, we developed and clinically introduced an AI-based fully automated brachytherapy treatment planning to substantially speed up the process. However, implant reconstruction for treatment planning is still manually performed and takes about 75 min on average per fraction [1]. Depending on the patient anatomy, several applicator types and sizes may be used. This study develops a deep learning-based workflow for automated applicator reconstruction in MRI-based brachytherapy, ensuring quick and precise reconstruction for all clinically used applicators. Material/Methods: MRI scans of 181 treatment fraction plans from 100 previously treated patients were included. The manually reconstructed applicator as was clinically used served as ground-truth. The Venezia applicator (Elekta AB, Sweden) was used in 73% and the Geneva applicator (Elekta AB, Sweden) in 27% of plans, with various ring/ovoid and intrauterine tube size combinations. The data was split into a training set with 145 plans from 80 patients (105 Venezia, 40 Geneva) and a test set with 36 plans from 20 patients (27 Venezia, 9 Geneva). The developed automated applicator reconstruction workflow uses an nnU-net [2] for segmentation and Elastix [3] for registration, as schematically described in Fig. 1. nnU-Net was trained for five configurations (‘2D’, ‘3D_lowres’, ‘3D_fullres’, ‘3D_cascade’ and ‘Ensemble’ [2]) to select the best. The workflow’s performance was evaluated by comparing automatically and manually reconstructed applicators through the Dice Similarity Coefficient (DSC) and dwell position errors (DPE, i.e., Euclidean distance between automatically and manually reconstructed dwell positions). Automated reconstruction time was recorded.