S1617
Physics - Autosegmentation
ESTRO 2026
segmentation achieved a Dice score of 0.469 ± 0.226, indicating greater anatomical variability.
accurate auto-contouring of thoracic OARs and GTV, substantially reducing manual effort. Its performance demonstrates clinical feasibility for integration into radiotherapy planning workflows, offering consistent delineations and supporting adaptive, AI-assisted treatment planning in thoracic oncology. Overall, the results confirm the effectiveness of leveraging frozen MedSAM features, with future work focusing on adaptive fine-tuning for improved tumor delineation. References: [1] Ma, Jun and Yang, Zongxin and Kim, Sumin and Chen, Bihui and Baharoon, Mohammed and Fallahpour, Adibvafa and Asakereh, Reza and Lyu, Hongwei and Wang, Bo, Medsam2: Segment anything in 3d medical images and videos, arXiv preprint arXiv:2504.03600, 2025[2] Çiçek Ö, Abdulkadir A, Lienkamp SS, Brox T, Ronneberger O. 3D U-Net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2016: 19th International Conference, Athens, Greece, October 17- 21, 2016, Proceedings, Part II 19 Springer; 2016. p. 424–432 Keywords: Thorax, Auto-contouring, Deep learning Digital Poster 5023 Internal Validation of a Deep Learning Auto- Segmentation for Thoracic Radiotherapy Planning CT Imaging Norina Predescu 1 , Monica E Chirilă 2 , Saad U Akram 1 , Jarkko Niemelä 1 , Jani Pehkonen 1 1 R&D Department, MVision AI Oy, Helsinki, Finland. 2 Radiation Oncology, Amethyst Radiotherapy Centre, Cluj-Napoca, Romania Purpose/Objective: Cardiovascular disease and cancer remain among the leading causes of mortality [1]. Long-term impairment of cardiac function represents a concern during the irradiation of thoracic malignancies Dose–response relationships have been investigated for various cardiac substructures; however, additional evidence is still required to strengthen these findings [2, 3]. This study evaluates the performance of a commercial CE- marked deep learning (DL)–based automatic segmentation device (Contour+TM) developed for radiotherapy treatment planning. The objective was to assess its accuracy across major cardiovascular and thoracoabdominal structures using clinically representative CT datasets. Material/Methods: A total of 118 thoracic radiotherapy planning CT scans were retrospectively collected. Ten (10) cardiac regions of interest (ROIs) were contoured according to international contouring guidelines by a team of
The model demonstrated robust geometric accuracy for well-defined organs and clinically acceptable boundary precision across thoracic structures. Test case sample slices are in the figure below.
Conclusion: The MedSAM+3D-UNet model enables fast and
Made with FlippingBook - Share PDF online