Radiation dose calculation in 3D heterogeneous media using artificial neural networks

Area of Science:

• Medical Physics
• Computational Dosimetry
• Machine Learning in Radiation Therapy

Background:

• Accurate dose calculation is crucial for external beam radiotherapy (EBRT) treatment planning.
• Current dose calculation methods, Monte Carlo (MC) and deterministic approaches, have limitations in speed and accuracy, especially near heterogeneities.
• Neural networks offer a potential solution for rapid and accurate dose distribution prediction.

Purpose of the Study:

• To develop a framework for training machine learning models to directly calculate radiation dose in 3D heterogeneous media.
• To enable faster and more accurate dose calculations for clinical treatment planning.

Main Methods:

• A novel framework for training machine learning models using spatially local information.
• Generation of 3D heterogeneous geometries using simplex noise for MC simulations.
• Development of precalculated data channels encoding beam parameters and contextual information for model input.

Main Results:

• A fully connected neural network model successfully reproduced MC dose distributions with high accuracy (94.7% gamma index pass rate, 1.45% average error).
• The model demonstrated generalization capabilities by accurately calculating dose in an unseen patient CT image.
• The framework enables rapid dose calculation, significantly faster than traditional MC methods.

Conclusions:

• A novel method for generating training data for radiation dosimetry models was introduced.
• The framework and preprocessing steps allow even simple models to achieve accurate dose distributions for EBRT.
• The demonstrated generalization indicates the potential for clinical application in predicting therapeutic doses.