ESTRO 2025 - Abstract Book

S3454

Physics - Machine learning models and clinical applications

ESTRO 2025

Conclusion: This study demonstrates the potential of DRL-based adaptive fractionation to spare OARs. Further research should focus on applying this approach to more realistic 3D scenarios and more complex plan variations.

Keywords: Adaptive fractionation, Reinforcement learning

References: [1] Haarnoja, Tuomas, et al. "Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor." International conference on machine learning . PMLR, 2018. [2] Huang, Shengyi, et al. "Cleanrl: High-quality single-file implementations of deep reinforcement learning algorithms." Journal of Machine Learning Research 23.274 (2022): 1-18. [3] Haas, Y. Pérez, et al. "Adaptive fractionation at the MR-linac." Physics in Medicine & Biology 68.3 (2023): 035003.

4519

Digital Poster Early Error Detection in MR-Guided Online Adaptive Radiotherapy: A Peer Review Approach Siqiu Wang, Christopher Kabat, Justin Visak, Yesenia Gonzalez, Ruiqi Li, Sean Domal, Ying Zhang, Steve Jiang, Zohaib Iqbal, Tsuicheng Chiu, Fan Chi Su, Mu-Han Lin Radiation Oncology, The University of Texas Southwestern Medical Center, Dallas, USA Purpose/Objective: Magnetic resonance (MR)-guided adaptive radiotherapy represents a significant advancement in personalized radiation oncology. However, the relatively recent adoption of this technology, coupled with the complex treatment planning process, underscores the critical importance of robust quality assurance measures, essential not only to mitigate the risk of medical errors but also to optimize the efficiency and accuracy of the online adaptive workflow. This study introduces and implements a planning peer review process incorporating a comprehensive checklist to facilitate upstream plan evaluation before physician approval, thereby preventing errors propagating downstream in the MR-guided online adaptive radiotherapy workflow. Material/Methods: The peer review process is conducted following the completion of the initial pre-plan and prior to physician approval, with a primary focus on identifying upstream errors to enable timely corrections before degradation of the adaptive workflow downstream. This process is led by the treatment planner and the chart checker, with open participation from the MR-guided radiotherapy physics group. The review was facilitated using an integrated checklist developed within an in-house scripted plan-check environment, with the checklist items selected based on clinical error analyses conducted by a team of experienced physicists. Over a three-month implementation period, 99 cases underwent peer review and were categorized into three groups: “proceed as is”, “minor revisions”, and “major revisions”, depending on the type and number of identified issues. Results: Among the 99 cases reviewed, a total of 220 issues were discovered, with recommendations being “proceed as is” for 13 cases (13.1%), “minor revisions” for 75 cases (75.8%), and “major revisions” for 11 cases (11.1%). The number of categories of issues identified are illustrated in Figure 1, while the session duration over time is presented in Figure 2.The highest percentage of issues were found in the category Structures (29.1%), followed by IMRT Constraints (18.2%) and Physician Intent (14.5%). Over 3 months of implementation, the average review session length by month reduced from 37.6 min to 23.7 min, suggesting improvement in efficiency.