S1618
Physics - Autosegmentation
ESTRO 2026
medically trained professionals and underwent at least 2 rounds of peer reviews, to ensure accuracy. Performance was measured using the Dice Similarity Coefficient and Hausdorff Distance (HD). Mean and median values were calculated for each ROI to represent both central tendency and data stability. Results: The mean Dice coefficients were situated between 0.57 for smaller, fine-detailed structures (A_LAD) to 0.96 for larger vessels (Heart). The mean Hausdorff 95% distances varied between 4.0 mm (A_Aorta) and 17.1 mm (A_LAD), indicating submillimetric to low- millimetric precision for most structures. The Ventricle_L exhibited high segmentation performance, with a mean Dice coefficient of approximately 0.87, mean Hausdorff 95% distance below 7 mm, suggesting that the DL-model achieved strong spatial agreement and boundary precision for this structure. Table 1 presents the results of the analysis.Table 1. Count (n) and similarity metrics (Mean Dice, Median Dice, Mean Hausdorff 95% and Median Hausdorff) for each analyses ROI.
These findings suggest that the developed model can significantly streamline radiotherapy planning or research projects by delivering accurate, automated structure delineations, thereby reducing manual workload and improving workflow consistency. Further real-world data is needed. References: 1. GBD 2019 Diseases and Injuries Collaborators. Global burden of 369 diseases and injuries in 204 countries and territories, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019 [published correction appears in Lancet. 2020 Nov 14;396(10262):1562]. Lancet. 2020;396(10258):1204- 1222. doi:10.1016/S0140-6736(20)30925-92. Banfill K, Giuliani M, Aznar M, et al. Cardiac Toxicity of Thoracic Radiotherapy: Existing Evidence and Future Directions. J Thorac Oncol. 2021;16(2):216-227. doi:10.1016/j.jtho.2020.11.0023. Cicchetti A et al. Cardiac Survival Limiting Toxicity in Locally Advanced Lung Cancer Patients Treated with RT.Radiotherapy and Oncology. Volume 194 Supplement 1, May 2024 Keywords: deep learning, heart substructures, contouring Guideline-Compliant AI Auto-Segmentation for Pelvic Lymph Nodes in Gynecological and Prostate Cancer: Multicenter Validation Norina Predescu 1 , Join Y Luh 2 , Robert A Zlotecki 3 , Gabriel Axelrud 4 , Jarkko Niemelä 1 , Jani Pehkonen 1 1 R&D Department, MVision AI Oy, Helsinki, Finland. 2 Department of Radiation Oncology, Dr. Russel Pardoe Radiation Oncology Center, Eureka, CA, USA. 3 Department of Radiation Medicine, Medical University of South Carolina, Charleston, SC, USA. 4 Department of Radiation Oncology, Joe Arrington Cancer Research and Treatment Center, Covenant Health, Lubbock, TX, USA Purpose/Objective: Pelvic lymph node (LN) delineation forms the foundation of elective nodal irradiation (ENI) in gynecological and prostate cancer treatment planning, Digital Poster 5053 but remains a time-consuming and variable step. Accurate, consistent LN region contouring impacts treatment quality and clinical throughput.This validation study demonstrates how AI-powered auto- segmentation delivers clinically acceptable guideline- compliant pelvic LN contours across diverse patient populations. Material/Methods: A test set of 50 pelvic radiotherapy CT scans (25/gender) were used as test dataset. Four (4) scans originated from public datasets (TCIA)1 and 46 from partner clinics in Finland, Singapore, Germany, Estonia
Conclusion: The results show that the DL model achieves high segmentation accuracy for major cardiovascular structures on radiotherapy planning CT images. Large and mid-sized vessels and chambers (aorta, ventricles, atria, and vena cava) were segmented with good overlap and boundary conformity. The lower similarity metrics observed for A_LAD reflect the small diameter, elongated geometry, and proximity to complex surrounding tissues, which challenge both contrast differentiation and voxel-level prediction accuracy in CT imaging. While these results are typical for fine structures, they suggest that close manual review and potential refinement are necessary before clinical use.
Made with FlippingBook - Share PDF online