DeepBeam: a machine learning framework for tuning the primary electron beam of the PRIMO Monte Carlo software

Area of Science:

• Medical Physics
• Computational Physics
• Machine Learning Applications

Background:

• Monte Carlo (MC) simulations are crucial for accurate dose delivery in radiotherapy.
• Simulating megavolt photon beams requires precise electron beam characteristics, which are often unknown for medical accelerators.
• Accurate electron beam modeling is essential for reliable MC simulations.

Purpose of the Study:

• To develop a flexible framework using machine learning regression models.
• To estimate primary electron beam parameters for medical linear accelerator MC simulators.
• To utilize measured reference dose profiles in a water phantom for parameter estimation.

Main Methods:

• Developed two machine learning regression models: Principal Component Analysis (PCA) with Support Vector Regressors (SVR) and a deep learning model.
• PCA model extracts key features from dose profiles for SVR-based parameter estimation.
• Deep learning model uses encoders and fully connected layers for direct parameter regression from dose profiles.
• An optimization procedure refines parameter estimates by minimizing differences between measured and reconstructed dose profiles.

Main Results:

• Evaluated models using measured dose profiles from a Varian 2300 C/D accelerator and simulated datasets.
• The two-stage regression and reconstruction-minimization procedure yielded consistent electron beam parameter estimates.
• Achieved very good agreement between measured and simulated photon beam dose profiles.

Conclusions:

• The proposed framework provides a customizable and applicable tool for tuning virtual electron beams in MC simulators.
• The developed models enhance the accuracy of dose delivery simulations in radiotherapy.
• Open-source code, data, and procedures are available for broader application and research.