S140
Brachytherapy - Physics
ESTRO 2026
AIC is 0.96 and HD is 0.68mm, compared to 0.84 and 3.14mm for C-C vs AIC-AIC.Comparing AI-corrected to clinical delineations, DVI differences of a physical dose up to 0.75 Gy are observed for the bladder (see Figure 1), despite the DICE score for the bladder being 0.97 (HD 0.53mm). Overall, dosimetric differences are primarily linked to differences in bowel, sigmoid, and bladder delineations (see Figure 2). DVI differences purely as a result of AI-assisted needle reconstruction (C-C vs AIC-C) are small (<0.1Gy), only for the GTVRES differences up to 0.2Gy are observed.
Poster Discussion 4237 Evaluation of an AI-assisted scan-to-plan workflow for MRI-guided cervical HDR brachytherapy Renzo J. Scholman 1,2 , Vangelis Kostoulas 3 , Ellen M. Kerkhof 3 , Lavinia A.L. Verhagen 3 , Laura A. Velema 3 , Carien Creutzberg 3 , Peter A.N. Bosman 1,2 , Tanja Alderliesten 3 1 Evolutionary Intelligence group, Centrum Wiskunde & Informatica, Amsterdam, Netherlands. 2 Algorithmics Group, Delft University of Technology, Delft, Netherlands. 3 Dept. of Radiation Oncology, Leiden University Medical Center, Leiden, Netherlands Purpose/Objective: MRI-guided HDR brachytherapy for cervical cancer is a complex, time-critical procedure requiring, amongst others, delineation, implant reconstruction, and treatment planning. We present a novel AI-assisted scan-to-plan workflow, where both organs at risk (OARs) delineation and needle reconstruction are semi-automated, combined with automatic treatment plan optimization. We evaluate the impact of the (semi-)automated delineation and reconstruction part
on optimized plans. Material/Methods:
Retrospective data from 108 patients were used to develop the AI-assisted workflow. OARs delineation and needle reconstruction are performed using two 3D nnU-Net models [1] trained on MRIs with clinical delineations and reconstructions. Needles are automatically reconstructed by using an algorithm designed to mitigate reconstruction errors [2]. Treatment plans are generated using BRIGHT [3], following the EMBRACE-II protocol of 4 fractions of 7Gy [4]. Ten unseen patient cases are used for evaluation, with 2-6 interstitial needles each. We determine delineation and reconstruction correction times, and DICE score and Hausdorff distance (HD). Furthermore, for each case, clinical (C), AI, and AI-corrected (AIC) versions of reconstructions R and delineations D (denoted as R-D configurations) are considered. We compare BRIGHT-optimized treatment plans that adhere to all EMBRACE-II aims between various R-D combinations to analyze the dosimetric difference of dose-volume indices (DVIs) by, for each patient, using the dwell times obtained for each configuration in the other configurations. Results: Clinical workflow is timed for nine applications in five separate patients, revealing 27±9 minutes for delineation and 11±5 minutes for needle reconstruction. AI-assisted workflow is timed for ten delineations and four needle reconstructions, revealing 16±8 minutes and 5±3 minutes respectively.Most AI outputs required only minor corrections. Median DICE score between AI-AI and AIC-
Made with FlippingBook - Share PDF online