S1558
Physics - Autosegmentation
ESTRO 2026
Results:
Daniel Höfler 1 , Philipp Schubert 1 , Natalia Belmas 1 , Alina Depardon 1 , Ana-Maria Dubina 1 , Ahmed Gomaa 1 , Matthias May 2 , Benjamin Frey 1 , Udo Gaipl 1 , Christoph Bert 1 , Annette Schwarz 1 , Stefanie Corradini 1 , Florian Putz 1 1 Department of Radiation Oncology, Universitätsklinikum Erlangen, Friedrich-Alexander- Universität Erlangen-Nürnberg, Erlangen, Germany. 2 Department of Radiology, Universitätsklinikum Erlangen, Friedrich-Alexander-Universität Erlangen- Nürnberg, Erlangen, Germany Purpose/Objective: Promptable auto-segmentation models represent a novel paradigm for auto-contouring, enabling interactive user-guided segmentation, yet their clinical utility in radiotherapy planning remains unclear. We therefore conduct a real-world evaluation of nnInteractive, a recent top-performing 3D U-Net– based promptable segmentation model, to assess segmentation quality and efficiency for gross tumor volume (GTV) delineation in radiation therapy. Material/Methods: Three radiation oncologists delineated 25 GTVs that were randomly selected from clinical routine using two approaches: manual contouring using standard clinical contouring tools (syngo.via, Siemens Healthineers) followed by promptable auto-segmentation with nnInteractive (Isensee et al. 2025,implemented in 3DSlicer open-source software). Real-world time- saving benefits of promptable auto-contouring with nnInteractive vs. standard manual delineation were measured using stop watches. GTV auto-segmentation quality was assessed by determining geometric accuracy in reference to corresponding manual delineations. Moreover, a blinded expert rating was performed, including three independent radiation oncologists (five-point scale, 1 corresponding to the lowest and 5 to the highest rating). Results: GTVs spanned four anatomic regions (brain, head/neck, thorax, pelvis) and underlying imaging modality was MRI in 80% (20/25) and CT in 20% (5/25) of cases. Radiation oncologists performed interactive auto-contouring with Scribble prompts in 88% (22/25) and Lasso prompts in 52% (13/25) of GTVs, whereas point prompts were used in only one case. Average manual GTV contouring time was 5:05 min (range, 0:40 – 34:47 min), whereas average interactive auto- contouring with nnInteractive took 0:55 min (0:15 – 04:57 min) corresponding to a time saving benefit of 4:10 min (range, -0:04 – 29:50 min) per GTV. Time saving benefit was independent of tumor volume (r = 0.09, p = 0.661) and lesion sphericity (r = -0.05, p = 0.800). Interactive GTV auto-segmentations achieved a mean dice similarity score (DSC) of 0.86, a mean surface DSC of 0.91 (1 mm tolerance) and a mean
SAM-FT achieved a mean Dice of 0.770 ± 0.126, comparable to nnU-Net (0.772 ± 0.098) and higher than SAM-Med3D (0.507 ± 0.213). SAM-FT showed superior performance for larger soft-tissue structures such as the oral cavity and larynx, while nnU-Net maintained an advantage on smaller, low-contrast organs such as the cochlea and esophagus. The fine- tuned SAM-FT model achieved a balanced performance across organ sizes, with inference time per volume reduced from approximately 45 s (nnU- Net) to 8 s, representing a more than fivefold improvement in speed. Conclusion: Fine-tuning a foundation model for domain-specific 3D medical segmentation substantially improves both quantitative performance and computational efficiency. SAM-FT’s transformer-based contextual encoding and adaptive prompting enhance geometric precision, reduce inter-observer variability, and accelerate contour generation. These findings support the clinical feasibility of foundation-model adaptation for real-time, adaptive radiotherapy workflows. References: 1.Wang, H., Guo, S., Ye, J., Deng, Z., Cheng, J., Li, T., Chen, J., Su, Y., Huang, Z., Shen, Y., Fu, B., Zhang, S., He, J., & Qiao, Y. (2023). SAM-Med3D: Towards general- purpose segmentation models for volumetric medical images.arXiv preprint arXiv:2310.151612.Isensee, F., Jaeger, P. F., Kohl, S. A. A., Petersen, J., & Maier-Hein, K. H. (2021). nnU-Net: A self-configuring method for deep learning-based biomedical image segmentation.Nature Methods, 18(2), 203–211 Keywords: head and neck, Foundation Model Fine- Tuning Digital Poster 1739 Automatic but interactive: A real-world evaluation of promptable AI auto-segmentation for GTV delineation (PROMPTO)
Made with FlippingBook - Share PDF online