S2988
Invited Speaker
ESTRO 2026
new challenges compared to conventional, deterministic systems. This talk will address the transition
be safely adjusted by the user. We will examine key aspects of MLC modelling, including leaf-tip geometry and tongue-and-groove behaviour. Pre-commissioned data Modern linac manufacturers now provide “golden beam data,” which, given the improved reproducibility of current linac production, may in some cases offer better agreement with TPS modelling than user-collected datasets.
from preconfigured to pre-trained systems and the associated risks, emphasizing that AI tools must be commissioned and validated differently. Rather than focusing solely on aggregate performance metrics, commissioning should evaluate model behavior under varying conditions, including data dependence, robustness, and potential failure modes, while clearly defining the intended clinical scope of use. A central theme is the necessity of local validation, as model performance can vary significantly across institutions due to domain shift in imaging protocols, patient populations, and clinical practice. Building on this, a risk-based framework for evaluation will be presented, linking model performance to potential clinical impact and safety thresholds. The talk will further outline a lifecycle-based deployment strategy, including prospective validation in shadow mode, phased clinical introduction, and continuous monitoring to detect performance drift and guide re-validation and model updates. Finally, the importance of a human-in-the-loop approach will be highlighted, ensuring that clinicians remain central to safe and effective AI integration. Overall, this presentation aims to provide a practical framework for the safe commissioning, validation, and clinical implementation of AI-based tools in radiation oncology. 5276 Linear accelerators: Ensuring safety beyond the vendor's commissioning Eduard Gershkevitsh Radiotherapy, North Estonia Medical Centre, Tallinn, Estonia As a linear accelerators generated beams and mechanics becoming more and more reproducible, there is a general vendor push towards a pre- configured TPS beam models. MLC modelling in TPS have been attributed to a poor dosimetric audit results and mismatch between planned and delivered dose distribution*. This trend could not only accelerate the acceptance and commissioning time of the linac, but could also result in a more accurate and safer radiotherapy. The presentation will also discuss the advantages and disadvantages of vendor led commissioning as well as need to put more efforts towards system validation. References: *Glenn et al. "Photon beam modelling variations predict errors in IMRT dosimetry audits". Radiotherapy and Oncology 66: 8-14, 2022
In addition to TPS configurations, phantom manufacturers increasingly recommend
electron-density overrides for their phantoms. With the growing implementation of dose-to-medium algorithms rather than dose-to-water–based approaches, these recommendations merit renewed scrutiny. PDCA Finally, we will discuss the value of class solutions and the benefits of maintaining a database of quality-assurance results to continually refine agreement between calculated and measured dose. This need is amplified as treatment planners create increasingly complex plans that push both the TPS model and linac hardware toward their operational limits. References: Hernandez V, Vera-Sánchez JA, Vieillevigne L, Saez J. Commissioning of the tongue-and-groove modelling in treatment planning systems: from static fields to VMAT treatments. Phys Med Biol. 2017 Aug 1;62(16):6688- 6707. doi: 10.1088/1361-6560/aa7b1a. Kerns JR, Stingo F, Followill DS, et al. Treatment Planning System Calculation Errors Are Present in Most Imaging and Radiation Oncology Core-Houston Phantom Failures. Int J Radiat Oncol Biol Phys. 2017
Aug 1;98(5):1197-1203. doi: 10.1016/j.ijrobp.2017.03.049.
Schuring D, Westendorp H, van der Bijl E, et al. NCS35: Quality assurance of Treatment Planning Systems, Delft, The Netherlands: Netherlands Commission on Radiation Dosimetry, 2022 doi: 10.25030/ncs-035
5275 Artificial Intelligence in radiotherapy: Commissioning and validation of AI-based tools Gerd Heilemann Departement of Radiation Oncology, Medical University of Vienna, Vienna, Austria Artificial intelligence (AI) is rapidly entering clinical radiotherapy workflows, particularly through vendor- provided, pre-trained models for tasks such as segmentation, treatment planning, and decision support. While these tools promise substantial efficiency gains, they also introduce fundamentally
Made with FlippingBook - Share PDF online