ESTRO 2025 - Abstract Book

S2563

Physics - Autosegmentation

ESTRO 2025

Conclusion: The integration of AI-powered tools into the radiation oncology workflow enables rapid and precise delineation, with an average time of approximately 1.5 minutes for liver and GTV contouring. This efficiency supports the feasibility of quickly assessing surgical resection versus SBRT based on remaining liver volume, facilitating informed and timely decisions in multidisciplinary team settings. Such AI tools are a promising addition to the future of radiation oncology and personalized liver metastasis treatment planning.

Keywords: AI-Powered Contouring, Future Liver Remnant

References: 1. Rodríguez, M.C.R., Chen-Zhao, X., Hernando, O., et al. (2024). SBRT-SG-01: Final results of a prospective multicenter study on stereotactic body radiotherapy for liver metastases. Clinical and Translational Oncology , 26, 1790–1797 2. Doi, H., Beppu, N., Kitajima, K., & Kuribayashi, K. (2018). Stereotactic Body Radiation Therapy for Liver Tumors: Current Status and Perspectives. Anticancer Research , 38(2), 591–599. 3. Mahadevan, A., Blanck, O., Lanciano, R., et al. (2018). Stereotactic Body Radiotherapy (SBRT) for liver metastasis – Clinical outcomes from the international multi-institutional RSSearch® Patient Registry. Radiation Oncology , 13, 26.

4458

Digital Poster Robustness of automatic organ segmentation on various image contrasts from photon-counting CT Sebastian Baum, Patrick Wohlfahrt Varian, Cancer Therapy Imaging, Siemens Healthineers, Forchheim, Germany Purpose/Objective: Accurate and precise segmentation of anatomical structures is pivotal for high-conformal treatment techniques. Automatic organ segmentation facilitates this task and is commonly applied on conventional 120kVp single-energy CT (SECT120). This study assessed the robustness of automatic organ segmentation with respect to various image contrasts derived from photon-counting CT. Material/Methods: Virtual monoenergetic images (VMIs) ranging from 40keV to 190keV were reconstructed from photon-counting CT scans of 17 thoracic and 13 head-and-neck cancer patients acquired on NAEOTOM Alpha (Siemens Healthineers) using smooth (Qr40, Br40) and sharp (Br56, Br64, Bl64, Br76) kernels. Iterative metal artifact reduction (iMAR) was applied on head-and-neck cases with dental implants. Up to 150 anatomical structures were segmented with a two stage deep-learning architecture starting with anatomical landmark detection followed by organ segmentation in volumes-of-interest (syngo.via RT Image Suite, Siemens Healthineers). The impact of various image contrasts on landmark detection was assessed by the varying number of structures detected per patient. Segmentation conformity of organs contoured in all VMI reconstructions was quantified by Dice similarity coefficient (DSC), surface Dice (SD) with 1mm criterion and Hausdorff distance (HD) using a SECT120-equivalent image as reference (70keV VMI, Qr40). Results: The number of undetected anatomical landmarks increased with difference from reference VMI energy. The fewest organs were detected at 40keV (3.8 and 22 less organs on average for thoracic and head-and-neck cases, respectively). The segmentation conformity highly depends on VMI energy (Figure 1A) with median DSC and SD <0.9 and HD>3mm at 40keV. Low-energy VMI (higher contrast compared to reference, more sensitive to high-attenuating materials like metal) showed higher variability compared to high-energy VMI (lower contrast, reduced metal artifacts). Organ groups with high-contrast boundaries (lung, skeleton) were less influenced by energy as organ