Physics-driven learning of x-ray skin dose distribution in interventional procedures

Area of Science:

• Medical Physics
• Radiological Imaging
• Computational Science

Background:

• Image-guided X-ray procedures carry risks of deterministic and stochastic effects due to accumulated radiation doses.
• Current methods for estimating patient skin dose, relying on air kerma at the interventional reference point (IRP) and backscatter factors, struggle to accurately model complex photon-tissue interactions.
• Monte Carlo (MC) simulations are the gold standard for skin dose modeling but are computationally intensive, limiting their real-time application.

Purpose of the Study:

• To develop a computationally efficient method for accurate estimation of skin dose distribution in image-guided X-ray procedures.
• To combine the accuracy of MC simulations with the speed of learning-based methods for real-time radiation risk monitoring.
• To reduce the computational complexity associated with traditional skin dose estimation techniques.

Main Methods:

• A hybrid approach integrating MC simulations with deep learning, specifically convolutional neural networks (CNNs).
• Initial photon propagation is rapidly estimated using ray casting (RC) based on the Beer-Lambert law.
• A CNN is trained using MC simulation results as ground truth, mapping RC outputs and patient model data (anatomy, material properties) to predict dose deposition.

Main Results:

• 163 MC and RC simulations were performed across different voxel phantoms and anatomical regions (head, thorax, abdomen, pelvis).
• The trained CNN achieved skin dose estimation with an error below 10% for most test cases.
• Edge-preserving smoothing (EPS) was utilized to mitigate uncertainties in MC simulations, particularly with fewer primary photons.

Conclusions:

• The integration of deep neural networks and MC particle physics simulations offers a viable path to significantly decrease computational time for accurate skin dose estimation.
• The proposed method can generate dose distributions in under one second on high-end hardware and within two minutes on lower-cost systems.
• Execution time is independent of the number of primary photons, making the approach robust and scalable for various clinical and research settings.