S1594
Physics - Autosegmentation
ESTRO 2026
Gonzalez ‐ Jimenez A et al ., Medical Image Analysis 105, 2025. https://doi.org/10.1016/j.media.2025.103735 Keywords: Supervised training, data quality, data annotation
dataset models. For K-means, we divided the data into groups based on training set sizes, then selected the patient closest to the centroid of each group to build the model. ABAS performance was evaluated using geometric metrics, including Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), and mean distance-to-agreement (MDA), by comparing with reference contours on remaining patients to assess inter-patient performance. Differences in assumptions and cohort size were assessed using the Wilcoxon rank-sum test (p<0.05; 95%CI).
Digital Poster 4124
Feasibility of single institutional atlas-based inter- patient auto-segmentation for prostate MR-guided radiotherapy in a simulation-free workflow Thee Mateepithaktham 1 , Tanwiwat Jaikuna 1 , Tissana Prasartseree 1,2 , Kongpop Chukly 1 , Sumetha Fang 1 , Naphat Komenake 1 , Akrapol Suppasedtanon 1 , Paritt Wongtrakool 1 , Natthanicha Sauenram 1 , Naritsa Rotmuenwai 1 , Ratchapas Romrattaphan 1 , Wiwatchai Sittiwong 1,2 , Pongpop Tuntapakul 1 , Wajana Thaweerat 1 , Pitchayut Wongsuwan 1 , Marianne Aznar 3 , Eliana Vasquez Osorio 3 , Pittaya Dankulchai 1 1 Division of Radiation Oncology, Department of Radiology, Faculty of Medicine Siriraj Hospital, Mahidol University, Bangkok, Thailand. 2 Department of Radiation Oncology, University of Texas MD Anderson Cancer Center, Houston, Texas, USA. 3 Division of Cancer Sciences, School of Medical Sciences, Faculty of Biology, Medicine and Health, The University of Manchester, Christie NHS Foundation Trust Hospital, Manchester, United Kingdom Purpose/Objective: Simulation-free MR-guided radiotherapy (MRgRT) improves patient convenience, reduces registration errors, and shortens time-to-treatment. However, online adaptive contouring remains time-consuming because diagnostic and treatment-day positions differ. Upcoming AI auto-contouring for the MR male pelvis might assist, but its effectiveness is limited by the diversity of protocols and organ sets—for example, excluding the urethra. We propose an institutional atlas-based auto-segmentation (ABAS) method to generate initial contours on the first treatment day for MRgRT with an image-guided linear accelerator (MR- Linac) in routine clinical practice. Material/Methods: T2-MR images of 42 prostate cancer patients— recruited from a single institute—who underwent MR- guided radiotherapy, were analyzed in this study. Reference contours for prostate, seminal vesicles, bladder, urethra, and rectum were delineated on first- fraction T2-MR images by consensus of two experienced radiation oncologists (ROs). ABAS models were generated under two assumptions: 1) randomly sampled training sets and 2) K-means clustering using three factors—body volume, shape/texture, and position of OOIs (Figure 1). Training set sizes ranged from 10 to 22, increasing in increments of 4, while preserving data from the previous dataset for random
Results: The performance of ABAS-generated contours improved continuously as the number of patients included in the atlas library increased. The highest median (IQR) values of DSC, MDA, and HD95 were observed in the bladder from the atlas 22 k-means model (0.93±0.08, 0.10±0.17 cm, 0.93± 0.93 cm). The lower scores were found in small and difficult-to- identify structures, like urethras and seminal vesicles, with scores below 0.82, 0.05 cm, and 0.28 cm, respectively (Figure 2). However, no significant difference from the model created from a random training dataset.
Made with FlippingBook - Share PDF online