ESTRO 2025 - Abstract Book

S2544

Physics - Autosegmentation

ESTRO 2025

0.83±0.04 for the bladder, bowel, sigmoid colon and rectum, respectively. The corresponding HD95% values (mm) resulted: 4.32±4.20, 22.67±20.87, 14.66±17.66, and 5.64±2.30.

Conclusion: AI-based contouring showed promising results in MR-based gynaecological brachytherapy, demonstrating satisfactory agreement with manual reference contours, particularly for the bladder and rectum. However, the bowel proved to be the most challenging structure to delineate, with accuracy further affected by the complexity of MR imaging. Intra-observer variability was observed, especially for the bowel, warranting further exploration from a dosimetric perspective. Future research will involve a larger patient cohort, incorporate dosimetric analysis, and examine intra-observer variability across multiple physicians to better understand and optimize the role of AI in contouring workflows.

Keywords: autosegmentation, brachytherapy, MR-contouring

3786

Mini-Oral Validating an AI segmentation model for patients with prostate cancer: Does AI assisted delineation bias the clinician towards the AI segmentation? Henrik D Nissen 1 , Lars Fokdal 2,3 , Birgitte M Havelund 2 , Christine V Madsen 2 1 Department of Medical Physics, Vejle Hospital, University Hospital of Southern Denmark, Vejle, Denmark. 2 Department of Oncology, Vejle Hospital, University Hospital of Southern Denmark, Vejle, Denmark. 3 Department of Regional Health Research, University of Southern Denmark, Odense, Denmark Purpose/Objective: The goal of AI assisted delineation (AIaid) is to reduce inter-observer variation and ensure consistency in the delineation of tumours and organs-at-risk (OAR). Additionally, AIaid may shorten delineation time. However, concerns remain regarding the potential for AIaid to compromise individualization of delineations by biasing the clinician towards the AI segmentation. This study validated an AI segmentation model for the prostate, seminal vesicles, elective lymph nodes and OAR in prostate cancer (PCa) radiotherapy. The aim was to determine the model’s value from the clinician’s perspective in real-world, clinical use. Material/Methods: AI segmentations were generated using an in-house model. Three experienced oncologists delineated 30 PCa patients with and 30 without AI assistance as part of daily clinical work. In a subgroup of five patients, all clinicians performed AIaid along with a manual delineation. Delineation time was measured as the on-task time without interruptions. Clinical usefulness of AI segmentations was rated on a 4-point scale: No/minor/major corrections or reject. Additionally, the overall helpfulness of AI assistance was rated for each patient. Inter-observer variation was assessed in the five-patient subgroup. Results: AIaid reduced delineation time from a median of 42 (IQR 38–47) minutes to 24 (IQR 19.5–30) minutes. Scoring of individual structures revealed that most required no or few corrections. Elective lymph nodes (CTVn_E) and the bowel cavity were the volumes most often requiring major corrections. All participants found the AI model helpful (Figure 1). AIaid reduced inter-observer variation, as the AI segmentations were generally accepted. Corrections were predominantly local, with corrected inter-observer variation typically ranging from 4-6 mm. Larger variations (>1 cm) occurred when only some clinicians modified the AI segmentations. Large deviations mostly occur when the