S2408
Physics - Quality assurance and auditing
ESTRO 2026
Digital Poster 5032
analyzed. Each plan underwent standard QA measurement and an additional 5 mm offset measurement for independent verification. To assess robustness, delivery errors were simulated by perturbing the plan dose (e.g., underdose bands and grid misregistrations). The reconstructed dose maps were compared with ground - truth plan doses using 3%/2 mm gamma analysis, and full - map evaluations were performed against both measured and interpolated dose distributions. Reconstruction fidelity was further validated using the offset measurement data to assess prediction accuracy at unseen spatial positions. Results:
Development of complexity metric tolerances to support pre-treatment Patient Specific Quality Assurance (PSQA) reduction Kieran McAuley Medical Physics, St. Luke’s Radiation Oncology Network (SLRON), Dublin, Ireland Purpose/Objective: PSQA remains critical for verifying VMAT plan delivery. However, routine pre-treatment PSQA for all treatment plans is resource intensive, often providing limited additional information, and can potentially delay the start of treatment. This study aimed to develop and validate tolerances based on multiple complexity metrics to identify plans of increased complexity that warrant measurement, with the intention of reducing the overall PSQA workload in a busy hospital measuring approximately 1,000 PSQAs per year. Material/Methods: A total of 1,746 pre - treatment VMAT PSQAs were retrospectively analysed across a range of treatment sites. All plans were generated in the Monaco treatment planning system, delivered on Elekta Synergy linear accelerators, and measured using an ArcCHECK® detector array. Gamma analysis was performed using a global 3%/2mm criterion with a 10% dose threshold.An in - house Python software was developed to extract information from the DICOM plan files and calculate the complexity metrics, as shown below: Figure 1. Example of the Python software used to calculate complexity metrics and display the associated tolerance range to the userNine complexity metrics were investigated, with each respective output independently validated. As MLC modelling is a predominant source of PSQA deviations, six of the metrics specifically assessed MLC modelling and their mechanical parameters. Plans were categorised as ‘passing’ (> 95% gamma passing rate) or ‘failing’ (< 95% passing rate). Passing plans were used to establish tolerances and failing plans were used for validation of the tolerances. Results: Complexity metric tolerances were defined at the 3rd percentile in the direction of highest complexity for each metric’s distribution. Aperture Area Variability (AAV) and the average MLC speed were found to be
The proposed method achieved an average 3%/2 mm full-map gamma pass rate of 97.1%, outperforming conventional point gamma (93.0%) and interpolation (57.0%). The INR reconstruction accurately recovered structured delivery anomalies, such as horizontal underdose bands and misregistration shifts, which were not captured by standard QA. Independent offset measurements confirmed 100% gamma agreement with reconstructed predictions, demonstrating spatial accuracy and robustness. Conclusion: This prior-informed INR framework transforms conventional point-based QA into continuous, high- resolution verification without hardware modification. It enhances spatial fidelity, improves error detectability, and enables clinically interpretable dose distribution assessment, representing a practical advancement toward high-fidelity, AI-driven radiotherapy QA. References: [1] Li G et al., Evaluation of the ArcCHECK QA system for IMRT and VMAT verification, Phys Med, 2013, 29(3):295–303.[2] Chan M.F. et al., Using a novel dose QA tool to quantify the impact of systematic errors otherwise undetected by conventional QA methods: clinical head and neck case studies, Technol Cancer Res Treat, 2014, 13(1):57–67.[3] Sitzmann V et al., Implicit Neural Representations with Periodic Activation Functions, arXiv:2006.09661, 2020.[4] Tancik M et al., Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains, arXiv:2006.10739, 2020.[5] Ye S et al., Super-Resolution Biomedical Imaging via Reference-Free Statistical Implicit Neural Representation, Phys Med Biol, 2023, 68(20):acfdf1. Keywords: Super-resolution, dosimetry, deep learning
Made with FlippingBook - Share PDF online