Physics-informed machine learning for predicting MLC and gantry errors in VMAT: a feature engineering approach.

Area of Science:

• Medical Physics
• Machine Learning in Radiotherapy
• Radiation Oncology

Background:

• Volumetric Modulated Arc Therapy (VMAT) involves complex delivery dynamics.
• Accurate prediction of multileaf collimator (MLC) and gantry positional errors is crucial for VMAT quality assurance.
• Existing methods may not fully capture the intricate physics of VMAT delivery.

Purpose of the Study:

• To develop a physics-informed, feature-engineered machine learning (ML) framework for predicting MLC and gantry positional errors in VMAT.
• To introduce novel physics-based parameters to improve predictive accuracy.
• To compare the performance of different ML models for VMAT error prediction.

Main Methods:

• Utilized VMAT trajectory logs and DICOMRT plans from 32 TrueBeam linac treatments with HD120 MLC.
• Extracted delivery dynamics and engineered physics-based features (friction, gravity, MLC speed-normalized).
• Trained and optimized XGBoost, LightGBM, and deep neural networks (DNNs) using Optuna; evaluated feature importance with Spearman correlation, mutual information, and SHAP.

Main Results:

• Identified systematic discrepancies between DICOM-RT and trajectory log data (7-8.5% deviations).
• MLC speed was the dominant predictor (r_s=0.891); physics-driven features showed significant correlations.
• LightGBM and XGBoost achieved superior MLC error prediction (MAE: 0.0019 mm), reducing residual errors by 30%; DNNs performed less effectively.
• Gantry error prediction accuracy was lower (MAE: 0.012°-0.015°).

Conclusions:

• Domain knowledge integration in ML significantly enhances radiotherapy applications.
• Physics-based feature engineering achieved a 30% reduction in VMAT positional errors.
• Prioritizing feature space exploration alongside model optimization is recommended for VMAT quality assurance.