ESTRO 2025 - Abstract Book

S336

Brachytherapy - Physics

ESTRO 2025

methods in brachytherapy beyond the TG-43 formalism: Current status and recommendations for clinical implementation. Med. Phys. 2012, 39, 6208–6236

2611

Proffered Paper AI and the art of saving time: AI-based delineations in brachytherapy treatment planning. A paired prospective study. Maximilian Lukas Konrad 1,2 , Carsten Brink 1,2 , Irene Hazell 1 , Anja Ør Knudsen 3 , Gitte-Bettina Nyvang 3 , Trine Lambrecht Jørgensen 3 , Tine Schytte 3 , Ebbe Laugaard Lorenzen 1,2 1 Laboratory of Radiation Physics, Odense University Hospital, Odense, Denmark. 2 Department of Clinical Research, University of Southern Denmark, Odense, Denmark. 3 Department of Oncology, Odense University Hospital, Odense, Denmark Purpose/Objective: AI-based delineation tools are often motivated with more consistent delineations and delineation time reductions [1,2]. In this paired prospective trial, we investigated if using AI-based delineations followed by manual correction of experienced oncologists reduces delineation time in comparison to manual delineations only in PDR MR-guided brachytherapy for cervical cancer. The prospective nature of this study addresses the lack of in-clinic evidence regarding time-saving using AI-based delineation in radiotherapy treatment planning. Material/Methods: A nnU-Net[3] model was trained on 303 scans from 109 patients from our institution, using clinical data with automated curation. Regions of interest (ROIs) included bowel, bladder, rectum, sigmoid, HR-CTV, and GTV, delineated following EMBRACE II[4] guidelines. Delineations were done on T2-weighted MR scans. Delineation times from 28 individual patients with 54 PDR treatments were collected from December 2023 to November 2024. Each patient received two treatments as per clinical standards, with AI-based delineations provided to oncologists as an initial guide for the first treatment only. The second treatment was delineated manually without AI-based delineations. The primary endpoint was delineation time reduction with AI-based compared to manual delineations. Oncologists manually recorded delineation time, the delineating physician's name, and the perceived quality of the AI-based delineations on a 4-point grade (no correction needed to clinically unusable). Statistical analysis was performed using linear mixed models (LLM), with AI-based delineation time reduction as a fixed effect, and patient and delineating oncologists as random effects to account for inter-patient and inter oncologist variability. Bootstrap methods provided confidence intervals. Residuals were assessed for normality and homoscedasticity. Results: With the LMM a manual delineation time of 32.3 [95% CI: 30.4 34.2] minutes and a time reduction of 7.7 minutes [95% CI: 4.4-11.1] using AI were found. This corresponds to a relative time reduction of 25%. Raw delineation times can be seen in the boxplot in Figure 1.