S1707
Physics - Dose prediction/calculation, optimisation and applications for particle therapy planning
ESTRO 2026
Poster Discussion 834
A machine learning model for accurate spot-by- spot prediction delivery time in pencil beam scanning proton therapy Nicola Giuliani 1 , Giuseppe Magro 2 , Matteo Bagnalasta 3 , Arturs Meijers 4 , Nicola Bizzocchi 5 1 Treatment planning, MedAustron, Wiener Neustadt, Austria. 2 Medical Physics, CNAO, Pavia, Italy. 3 Radiation oncology, Universitätsspital Zürich, Zürich, Switzerland. 4 Medical Physics, PSI, Villigen, Switzerland. 5 Treatment planning, PSI, Villigen, Switzerland Purpose/Objective: Conventional Treatment Planning Systems (TPS) for Pencil Beam Scanning (PBS) proton therapy do not model spot-by-spot delivery times, limiting the ability to prospectively evaluate dose-rate distributions. This is a critical gap for research into the clinical effects of dose-rate and potential toxicities. This study aimed to develop and validate a machine learning model to accurately predict spot delivery duration using only pre-treatment planning parameters. Material/Methods: This research used a retrospective dataset from the Paul Scherrer Institute (PSI), consisting of log files from 218 delivered treatment beams, totaling over 200,000 spot measurements.A Python data pipeline was developed. It began with exploratory data analysis to understand the data’s structure and distributions. Data cleaning, consolidation, and statistical outlier removal using the Interquartile Range (IQR) method followed.Features engineering was key. Based on the delivery system’s physics, a novel feature, First Planned MU, was created to represent the machine’s layer-specific current calibration.A Random Forest Regressor was selected as the final model after comparing it to linear models and other tree-based ensembles. Hyperparameters were optimized using Bayesian Optimization with Mean Absolute Error (MAE) as the objective function.The model’s performance was evaluated on an unseen test set using standard statistical metrics (MAE, R ² ), feature importance analysis, and a stratified analysis of its uncertainty measure. Results: The final optimized Random Forest model demonstrated exceptional predictive power. On the unseen test data, the model achieved a Mean Absolute Error of 0.35 ms and a R ² score of 0.994, indicating it successfully explains over 99.4% of the variance in spot duration.
The feature importance analysis confirmed that Planned MU was the primary predictor (85.9% importance), as expected. Crucially, the domain- specific engineered feature, First Planned MU, was validated as the second most important predictor (8.8% importance), proving its high predictive value.
The most significant finding related to the model's uncertainty. A stratified analysis showed that the model's uncertainty was very low for common, short- duration spots (mean uncertainty ≈ 0.46 ms) but increased dramatically for rare, long-duration spots (mean uncertainty ≈ 38.84 ms). This directly correlates with the data scarcity in the training set, where over 99% of samples were for short-duration spots.
Conclusion: A machine learning model can accurately predict proton spot delivery times from planning data with high fidelity. The model provides a validated tool to enable pre-treatment dose-rate evaluation for research into its clinical effects. This work represents a proof of concept towards dose-rate-aware treatment planning in proton therapy. Keywords: dose-rate, prediction, machine learning Digital Poster 913 Validation of fully automated IMPT treatment planning for skull base chordomas Merle Huiskes 1 , Ida Coremans 1,2 , Koen Crama 1,2 , Steven Habraken 1,2 , Wens Kong 3 , Sebastiaan Breedveld 3 , Ben Heijmen 3 , Coen Rasch 1,2 , Eleftheria Astreinidou 1
Made with FlippingBook - Share PDF online