ESTRO 2025 - Abstract Book

S3658

Physics - Quality assurance and auditing

ESTRO 2025

References: 1. Taylor PA et al Prioritizing Clinical Trial Quality Assurance for Photons and Protons: A Failure Modes and Effects Analysis (FMEA) Comparison Radiother Oncol. 2023 May;182:109494. doi: 10.1016/j.radonc.2023.109494. 2. Huq, M.S., et al. (2016). Application of risk analysis methods to radiation therapy quality management: Report of Task Group 100 of the AAPM. Medical Physics, 43(7), 4209-4262. doi:10.1118/1.4947547.

3107

Poster Discussion Machine learning-based quality assurance tool for spot-scanning proton therapy: Comparing binary log files and DICOM treatment plans Sang Kyun Yoo 1 , Sridhar Yaddanapudi 2 , Bo Lu 2 , Ethan D. Stolen 3 , Siddhant Sen 4 , Byongsu Choi 2 , Jin Sung Kim 1 , Keith Furutani 2 , Chris Beltran 2 , James J. Sohn 3 1 Department of Radiation Oncology, Heavy Ion Therapy Research Institute, Yonsei University, Seoul, Korea, Republic of. 2 Department of Radiation Oncology 2, Mayo Clinic Florida, Jacksonville, USA. 3 Department of Radiation and Cellular Oncology, University of Chicago, Chicago, USA. 4 Department of Psychology, University of Illinois Urbana Champaign, Champaign, USA Purpose/Objective: Spot position accuracy in pencil beam scanning (PBS) proton therapy is crucial for treatment quality but faces challenges from beam monitoring and steering system uncertainties. Traditional quality assurance (QA) methods are time-consuming and may miss subtle delivery errors. This study develops and validates a machine learning (ML)- based tool designed to efficiently compare planned and delivered monitor units (MU) and spot positions, with the goal of improving patient-specific quality assurance (QA) workflows. Material/Methods: This study analyzed treatment plans and 40 delivery log files from a Hitachi PROBEAT-V proton system (20 files each from two gantry rooms), collected over one month for energies ranging from 100.7-124.8 MeV. Statistical analyses employed Levene’s test and Welch’s ANOVA to evaluate spot position errors across energy levels, treatment days, and rooms. ML models using XGBRegressor were developed with a novel feature engineering approach incorporating beam energy and gantry angle dependencies. The model training process included optimization of hyperparameters through grid search with 5-fold cross-validation. Model performance was evaluated using mean squared error (MSE), R² score, and Euclidean distance metrics. Feature importance was analyzed using SHAP (Shapley Additive exPlanations) values. Results: Statistical analyses revealed energy-dependent spot position variations, with significant deviations at 100.7, 103.6, 105.0, 110.5, and 119.7 MeV (P < 0.001). These variations strongly correlate with beam delivery parameters, particularly at energy extremes. The ML models achieved exceptional accuracy with R² scores of 0.999 for both coordinates. MSE values were 0.013 (x-coordinate) and 0.002 (y-coordinate), with an overall Euclidean distance error between 0.100 mm. Cross-validation across different treatment days showed consistent performance (coefficient of variation < 1.5%). Time efficiency analysis demonstrated a 75% reduction in QA analysis time compared to conventional methods. Finally, the SHAP analysis revealed that beam energy and gantry angle were the most influential features in predicting spot position deviations, accounting for 45% and 30% of the model’s decisions, respectively. Conclusion: This ML-based QA tool accurately predicts PBS spot delivery positions while significantly improving efficiency. The system’s ability to detect subtle delivery variations across different energies and treatment conditions offers the potential for real-time adaptive QA solutions. The tool’s robust performance across various treatment conditions and