Fast occupational upper-limb radiation dose prediction using machine learning and Monte Carlo simulation.

Area of Science:

• Medical Physics
• Radiological Sciences
• Machine Learning in Healthcare

Background:

• Interventional procedures pose occupational radiation risks to physicians' upper extremities.
• Current extremity dosimeters lack spatial information, real-time feedback, and have performance limitations.
• There is a need for advanced methods to accurately assess occupational radiation dose in interventional settings.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for estimating radiation dose at discrete upper-limb locations.
• To utilize Monte Carlo (MC)-derived data and procedure-specific parameters for dose prediction.
• To create a robust and scalable system for occupational dose assessment in interventional radiology.

Main Methods:

• Generated a dataset of 10,000 MC dose maps under diverse clinical and geometric conditions.
• Trained and validated various ML models, including deep neural networks and tree-based regressors, using five-fold cross-validation.
• Constructed an ensemble of the top three models to enhance prediction accuracy and robustness, evaluating performance using mean absolute and relative errors.

Main Results:

• The ensemble model achieved an average relative error of 3.69%, outperforming individual models.
• Consistent performance was observed across different anatomical regions and dose levels.
• Highest accuracy was noted for standard beam geometries, with minor discrepancies in extreme configurations.

Conclusions:

• ML-based extremity dose estimation is feasible for interventional environments.
• The proposed ensemble model offers a rapid (approx. 10 ms) and scalable alternative to full MC simulations.
• This approach enables near real-time prediction of upper-limb occupational dose, aiding radiation protection optimization.