S55
Brachytherapy - Gynaecology
ESTRO 2026
Proffered Paper 1590
70.1% vs 52.2% (p=0.79) for ISBT and ICBT, respectively.
Performance Evaluation of an AI-Driven Auto- Segmentation Tool for MR-Based Cervical Cancer Brachytherapy Planning Elaine Limkin 1 , Teo Kevin 2 , Sophie Espenel 1 , Isabelle Dumas 1 , Ines Chaffai-Nouri 1 , Alexandru Iancu 3 , Anne- Agathe Serre 4 , Frédéric Gasse 5 , Thomas Theodoridis 6 , Sami Romdhani 6 , Lhassa Macke 7 , Madalina-Liana Costea 7 , Nikos Paragios 8 , Charlotte Robert 1 1 Radiation Oncology Department, Institut Gustave Roussy, Villejuif, France. 2 Radiation Oncology Department, University of Pennsylvania, Philadelphia, USA. 3 Radiation Oncology Department, Institutul Oncologic “Prof. Dr. Ion Chiricu ță ”, Cluj-Napoca, Romania. 4 Radiation Oncology Department, Centre Leon Berard, Lyon, France. 5 Frederic Gassa frederic.gassa@lyon.unicancer.fr Medical Physics department, Centre Leon Berard, Centre Leon Berard, Lyon, France. 6 AI engineering, TheraPanacea, Paris, France. 7 Clinical Affairs, TheraPanacea, Paris, France. 8 CEO, TheraPanacea, Paris, France Purpose/Objective: Manual delineation of organs-at-risk (OARs) for cervical brachytherapy is labor intensive, time consuming, and subject to inter-observer variability. Deep-learning auto-segmentation offers a promising alternative that can accelerate treatment planning, thereby reducing the time between applicator implantation and start of treatment. We quantitatively assessed the clinical performance of an automatic- contouring solution for MR-based brachytherapy treatments and compared it with expert-generated contours. A clinical acceptability assessment was then performed with two radiation oncologists (ROs). Material/Methods: A 3D U-net [1] deep-learning architecture was used to
Figure 1. Kaplan-Meier plot of overall survival (A) and progression-free survival (B)Acute toxicities were mostly grade 1-2, with gastrointestinal (77%) and genitourinary (46%) events being the most common; grade ≥ 3 events were rare. Late toxicities were infrequent, with grade 1-2 gastrointestinal and genitourinary events in 17% and 8% of patients, respectively. No significant differences were observed between ISBT and ICBT for any toxicity. Univariate analysis showed ECOG status (p=0.23) and FIGO stage (p=0.16) to be significant factors. Multivariate analysis– including boost approach, ECOG status, and FIGO stage–identified no significant independent determinant for OS or PFS. Conclusion: Sequential adaptive EBRT boost combined with ICBT may result in comparable survival and toxicity outcomes to ISBT. However, our data should be interpreted with caution, as limitations include selection bias (with a higher proportion of FIGO IVA in the ISBT cohort), confounding bias and possible misclassification due to financial limitations that preclude optimal staging (96.55% versus 15.79% in ISBT and ICBT, respectively, had imaging done). Nevertheless, sequential adaptive EBRT boost combined with ICBT may be a viable alternative to ISBT when capability or resource is lacking. Keywords: Brachytherapy, Sequential Adaptive EBRT Boost References: 1Liu Z, Zhao Y, Li Y, et al. Imaging-guided brachytherapy for locally advanced cervical cancer: the main process and common techniques. Am J Cancer Res. 2020 Dec 1;10(12):4165-4177. PMID: 33414993; PMCID: PMC7783772.2Kim H, Kim YS, Joo JH, et al. Tumor Boost Using External Beam Radiation in Cervical Cancer Patients Unable to Receive Intracavitary Brachytherapy: Outcome From a Multicenter Retrospective Study (Korean Radiation Oncology Group 1419). Int J Gynecol Cancer. 2018 Feb;28(2):371-378. doi: 10.1097/IGC.0000000000001155. PMID: 29189448.
develop a contouring solution for four OARs systematically delineated for brachytherapy
treatments: bladder, rectum, sigmoid and bowel. The training was done using 327 MR T2 images with varied brachytherapy applicators (Vienna, Geneva, Ring, Venezia or personalized mold) from a single center in Europe, fully annotated by ROs. A test dataset of 42 patients (independent cohorts from two centers, 25 patients from the same European center as the training data and 17 patients from another US center) was used to assess model performance. The quantitative evaluation was done by computing Dice (DSC) and Hausdorff distance 95 (HD95) metrics between expert-based ground truth contours and the AI-based automatic contours. A subset of 18 patients from the test dataset was used to assess the clinical acceptability of the contours through a qualitative review of two expert ROs. A three-point scale was used to assess the clinical acceptability of each contour
Made with FlippingBook - Share PDF online