Vision Transformer model-based dose prediction and beam angle optimization for BNCT.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Accurate dose prediction is crucial for effective Boron Neutron Capture Therapy (BNCT) treatment planning.
• Current Monte Carlo (MC) simulations offer precision but are computationally intensive, limiting planning efficiency.
• Developing faster, accurate dose prediction methods is essential for optimizing BNCT plans.

Purpose of the Study:

• To develop an advanced neural network model for efficient and accurate prediction of BNCT dose distributions.
• To integrate this model with Bayesian optimization for selecting optimal beam angles.
• To enhance treatment planning by improving tumor dose delivery and minimizing dose to organs at risk.

Main Methods:

• A deep learning framework utilizing a 3D Vision Transformer (ViT) and Mamba module for dose prediction.
• Incorporation of a region of interest (ROI)-guided attention mechanism focusing on Gross Tumor Volume (GTV) and skin.
• Integration of predicted doses into a Bayesian optimization strategy for beam angle selection.

Main Results:

• The model achieved high accuracy in dose prediction, with Mean Absolute Error (MAE) below 0.6 Gy for GTV and below 0.15 Gy for skin.
• Gamma passing rates exceeded 90% (2 mm/2%) and 97% (3 mm/3%), indicating excellent agreement with MC simulations.
• Treatment optimization resulted in an average 1.8 Gy increase in GTV minimum dose without increasing Organ at Risk (OAR) maximum dose.

Conclusions:

• The proposed deep learning method provides accurate dose prediction and efficient optimization for BNCT.
• Results are validated against MC simulations, demonstrating its clinical potential.
• This approach could facilitate automated BNCT treatment planning and optimization.