Beyond the curve: a machine learning approach for wobbling proton beam therapy output prediction

초록

While analytical models for monitor unit prediction in wobbling proton beam therapy are well-established, they exhibit reduced accuracy in clinically extreme cases, such as large field sizes or seldom-used plan parameters. This study proposes a machine learning (ML) framework designed to act as a supportive verification tool for existing analytical models, enhancing predictive confidence across the full spectrum of treatment parameters. A pre-treatment quality assurance dataset comprising 2859 wobbling proton fields, collected between 2016 and 2024, was used to train and evaluate three ML architectures: random forest, gradient boosting machines (XGBoost), and neural networks. Input features included dosimetric and geometric variables such as distal ranges (D90, D10), spread-out Bragg peak (SOBP) length, and beam energy extracted from measurement sheets. The models were evaluated on their ability to provide reliable predictions for cases where the analytical model is known to exhibit significant deviations or fail entirely. The Machine Learning models demonstrated a strong ability to generalize from historical measurement data. For routine treatment parameters, ML predictions showed excellent agreement with the analytical fitting model. However, in extreme cases involving large field sizes and sparse data points, the ML models provided stable and reliable predictions where the analytical fitting model faltered. Among the test algorithms, XGBoost proved the most robust achieving a root mean squared percentage error of 2.79% compared to the analytical fitting model's 5.83%. Furthermore, the ML framework successfully generated monitor unit predictions for non-standard fields where the analytical fitting model provided no output. A Machine Learning framework serves as a powerful complementary tool for monitor unit prediction in wobbling proton beam therapy. By running in parallel with established semi-empirical models, it provides an essential safety layer, offering robust predictions for non-standard fields while independently confirming results for routine cases. This synergetic approach enhances the overall reliability and safety of the clinical quality assurance workflow without discarding the proven value of traditional analytical methods.

키워드