ESTRO 2025 - Abstract Book

S2484

Physics - Autosegmentation

ESTRO 2025

In the standard workflow, targets and organs at risk contours were propagated from reference scans to the scan of the day using deformable image registration and edited manually. In the AI-based workflow, AI delineations are manually imported and corrected by clinicians after auto-segmentation. Statistical significance in the difference in delineation time and total treatment time per fraction was tested with linear mixed models.

Results: The delineation time is shown for each fraction and group are summarized in Figure 2.

The AI-based workflow reduced the median online delineation time from 9.8 to 5.3 minutes (46 % time reduction) and the overall treatment time from 27.3 to 25.0 minutes (8 % time reduction). A statistically significant delineation time reduction (p-value < 0.001) and total treatment time reduction (p-value < 0.02) was found. Variance in delineation time increased during treatment for Group 1 but stayed consistent for Group 2. For one patient, the nnU-Net model performed sub-optimally, and the AI workflow was changed to a standard workflow for five fractions, while a patient-specific model was trained. Conclusion: Implementing AI delineation at the MRI-Linac is feasible and effectively reduces delineation time, with improved consistency in delineation time supporting better daily scheduling. In-house developed AI-based delineation tools allow training patient-specific models, yielding more precise and consistent delineations in the online workflow. These results highlight the potential of AI-based delineation to streamline and enhance the efficiency of adaptive radiotherapy.

Keywords: Prospective Study, AI, Timegain, MRI-Linac

References: [1] Lorenzen EL, Celik B, Sarup N, Dysager L, Christiansen RL, Bertelsen AS, et al. An open-source nnU-net algorithm for automatic segmentation of MRI scans in the male pelvis for adaptive radiotherapy. Front Oncol 2023. [2] Sarup N, Konrad M, Madsen MLMaL, Krogh SL, Hansen CR, Hazell I, et al. Implementing a simple, dynamic and controllable AI delineation tool in clinical settings. ICCR 2024 proceedings.