S2120
Physics - Inter-fraction motion management and daily adaptive radiotherapy
ESTRO 2026
an external testing dataset (18 patients, 129 fractions). The validation set was used to optimize the prompting strategy, while the test datasets provided independent evaluation of the segmentation accuracy.Subsequently, a dosimetric study including 20 patients was performed using a single fraction per patient. Daily treatment plans re-optimized using nnInteractive-generated GTV masks were compared with the clinical plans (Figure 2). The statistical significance of dosimetric deviations was tested with the Wilcoxon signed-rank test. Results: The most effective prompting setup used three prompts, i.e., propagated contours in sagittal, coronal and axial view with 1-voxel dilation. nnInteractive refinement improved contour propagation accuracy by max. 15% compared to affine image registration- based propagation. The Dice Score (DSC) improved from 0.73 to 0.82 in the internal testing dataset and from 0.64 to 0.77 in the external dataset. The improvement in comparison with the TM-based propagation was less pronounced (max. 5%), increasing DSC from 0.82 to 0.83 in the internal dataset and from 0.77 to 0.81 in the external dataset (Figure 1).No statistically significant differences, in GTV coverage and OARs DVH parameters, between ground truth and treatment plans were found, although there were a few outliers (2 out of 20) with unacceptable coverage drop. Statistically significant dosimetric differences were found only in D98 in PTV.
Conclusion: Daily refinement with nnInteractive allows fully automated GTV contour adaptation with superior accuracy, while maintaining dosimetric quality in the majority of patients. References: [1] Isensee F, Rokuss M, Krämer L, Dinkelacker S, Ravindran A, Stritzke F, et al. nnInteractive: Redefining 3D Promptable Segmentation. ArXiv 2025;abs/2503.08373.[2] Wei C, Eze C, Klaar R, Thorwarth D, Warda C, Taugner J, et al. Deep learning- based contour propagation in magnetic resonance imaging-guided radiotherapy of lung cancer patients. Phys Med Biol 2025;70:145018.
https://doi.org/10.1088/1361-6560/ade8d0. Keywords: Foundation models, promptable segmentation
Digital Poster 1087
Assessment of inter-observer delineation variability in pancreatic cancer patients undergoing online adaptive MR-guided SBRT Sara Poeta 1 , Akos Gulyban 1 , Zelda Paquier 1 , Madeline Michel 2 , Elisa Bodson 2 , Nicolas Jullian 2 , Robbe Vandenbegin 2 , Christelle Bouchart 2 1 Medical Physics Department, Institut Jules Bordet, Brussels, Belgium. 2 Radiation Oncology Department, Institut Jules Bordet, Brussels, Belgium
Made with FlippingBook - Share PDF online