S1598
Physics - Autosegmentation
ESTRO 2026
GTVGP segmentations. Model performance was quantified using the expected calibration error (ECE). Results:
Randomized Trial of 2-Fraction or 5-Fraction Magnetic Resonance Imaging–Guided Adaptive Prostate Radiation Therapy. Int J Radiat Oncol Biol Phys123, 773–782 (2025).3. Vercauteren, T., Pennec, X., Perchant, A. & Ayache, N. Diffeomorphic Demons Using ITK’s Finite Difference Solver Hierarchy. (2007).4. Stan Development Team. Stan User’s Guide, Version 2.37. https://mc-stan.org/docs/2_37/stan-users-guide- 2_37.pdf (2025). Keywords: Probabilistic segmentation, MR-linac, GTV boost Digital Poster Highlight 4280 Using audit log time stamps to assess editing time of deep learning generated contours Geert De Kerf 1,2 , Michaël Claessens 1 , Gabriele Balletti 3 , Jonas Söderberg 3 , Dirk Verellen 1,4 1 Radiotherapy, Iridium Netwerk, Antwerp, Belgium. 2 Faculty of Medicine and Health Sciences, University of Antwerp, Antwerp, Belgium. 3 Machine Learning, RaySearch Laboratories, Stockholm, Sweden. 4 Centre for Oncological Research, University of Antwerp, Antwerp, Belgium Purpose/Objective: Deep learning–based contouring has been shown to reduce time requirements in clinical workflows at the task level [1]. To quantify time savings at the level of an organ, the added path length (APL) metric is commonly used, as it correlates with the time required for manual adjustments [2]. However, automatically generated contours can be modified using a variety of tools, questioning the correlation between APL and actual editing time. A new method, using audit log data extraction, is proposed to more accurately assess editing time for each individual contour. Material/Methods: In this study, we used audit log–based editing times from a large clinical dataset to benchmark against APL. Audit logs from RayStation (RaySearch Laboratories) were used to extract detailed information about which organ was edited, which tool was used, and the duration of each editing action. The extracted data were first validated against manually recorded editing times for 54 structures. Subsequently, the analysis was extended to 14,377 clinical organs at risk (OARs) that were initially segmented using a deep learning algorithm.An organ-specific analysis focused on the spinal cord, esophagus, heart, and left and right lungs, with at least 1,200 cases per organ. The correlation between editing time and APL was evaluated using the Pearson correlation coefficient. Results: The correlation between manual time recordings and audit log–based measurements was 0.98. Across the
The GTVGP segmentations generated for all fractions on the test set were well calibrated and aligned with GTVdaily, achieving an ECE of 0.05 (post-calibration), as shown in Figure 1. Example GTVGP probability maps are shown in Figure 2. Conclusion: A GP based on prostate contour correspondence can localize the GTV on MRdaily images. This could automate a subjective task while enabling probabilistic segmentation for applications in planning, dose reconstruction, biomarker quantification, and model explainability. References: 1. Poon, D. M. C. et al. Magnetic Resonance Imaging- guided Focal Boost to Intraprostatic Lesions Using External Beam Radiotherapy for Localized Prostate Cancer: A Systematic Review and Meta-analysis. Eur. Urol.6, 116–127 (2022).2. Cooper, S. et al. HERMES:
Made with FlippingBook - Share PDF online