ESTRO 2025 - Abstract Book

S1773

Clinical – Upper GI

ESTRO 2025

[95% CI: -5.01 to 13.34]) and 6 months (MD -11.11 [95% CI: -28.31 to 6.08]. The most significant decline was seen in emotional and social functioning. Fatigue worsened at 6 months. Whilst pain initially worsened at 3 months (MD 13.89), then improved by 6 months (MD 2.22). The 1-year overall survival rate was 68.18% (95% CI: 16.28% to 92.24%). Conclusion: Treatment of patients with unresectable pancreatic cancer on MR-Linac is feasible and well tolerated in the Australian public hospital setting. Further prospective studies are required to leverage the potential of MR-guided adaptive radiotherapy in this cohort of patients.

Keywords: Pancreas, MR-Linac, Quality of Life

1486

Digital Poster Evaluating efficacy of two AI auto-contouring algorithms for abdominal organs at risk (OAR) delineated in liver stereotactic body radiotherapy (SBRT) Matthew Lee 1,2 , Han Cheng 1 , Minxuan Yuan 1 , Glen Blackman 2 , Tom Richards 2 , Charles-Antoine Collins-Fekete 1 , Maria Hawkins 1,2 , Douglas Brand 1,2 1 Department of Medical Physics and Bioengineering, University College London, London, United Kingdom. 2 Department of Clinical Oncology, University College London Hospital NHS Foundation Trust, London, United Kingdom Purpose/Objective: During radiotherapy planning, clinicians contour the cancer target and OARs on patient cross-sectional imaging. This is predominantly done manually and can be time-consuming and subjective. AI models permit automated OAR segmentation, offering time-savings and greater reproducibility. Such commercial auto-contouring tools are being rapidly adopted for OAR contouring, although comparative data with humans is sparse in some sites (e.g. upper abdomen). Given the costs of such algorithms, interest has increased in open-source solutions. However, direct comparative performance against commercial solutions has not previously been reported. Our aim is to compare the performance of a commercial and open-source algorithm against human contours for upper abdominal OARs. Material/Methods: Patients treated with SBRT for hepatocellular carcinoma were identified from the local database between May 2020 and July 2023. 65 CT scans were pseudonymised and extracted in DICOM format. Manual segmentations by consultant clinical oncologists were obtained as ground-truth references. One open-source (algorithm A) and one commercial (algorithm B) algorithm was chosen based on OAR coverage and reported performance. Auto contoured organs included duodenum, stomach, left and right kidneys, spinal canal, and liver. Given potential variability in delineation of the stomach/duodenum border, a combined structure was also created. Outputted OARs were truncated cranio-caudally to match the human contour (humans typically only contour OARs near the target). Dice similarity coefficient (DSC) and 95 th percentile Hausdorff distance were calculated. Algorithm scores were compared by Wilcoxon Signed-Rank testing. Results: 51/65 (78.5%) of participants were male. Median age was 71 years (interquartile range 9.4). Tables 1 and 2 summarise the contouring performance of each algorithm. For both algorithms, performance was acceptable (DSC ≥ 0.8) on all organs besides duodenum, which performed better combined with the stomach; on visual inspection likely due to discrepancy in inter-structure transition point (figure 1). Algorithm B performed better in the kidneys, and algorithm A in the spinal canal due to algorithm B contours terminating at L1/L2 (figure 2). However, given the small absolute value differences, this is unlikely clinically significant.