ESTRO 2025 - Abstract Book

S3461

Physics - Optimisation, algorithms and applications for ion beam treatment planning

ESTRO 2025

These were used in four machine learning models—multivariate linear regression, random forest, XGBoost, and support vector machine (SVM)—to predict proton range shifts using machine learning. Model training and validation were performed by splitting the static spots in a 5:2 ratio for training and validation, while the accumulated scanned spots were exclusively used for validation. Model performance was evaluated using the root mean square error (RMSE).

Results: Temporal features alone provided accurate predictions, with RMSE values between 2 and 4 mm for the static spots, consistent with prior models [1] (Figure 2). Spectral information, including energy and higher-order image features, yielded poor prediction performance and had RMSE values exceeding 5 mm. Combined feature sets, which incorporated both temporal and spectral information, showed similar results to temporal features alone, indicating no benefit from spectral information in this PGT system.