Proton dose calculation with transformer: Transforming spot map to dose.

Area of Science:

• Medical Physics
• Computational Biology
• Radiotherapy

Background:

• Conventional proton dose calculation methods present a trade-off between computational efficiency and accuracy.
• Monte Carlo (MC) simulations are accurate but time-intensive, while analytical methods are faster but less precise.
• There is a clinical need for improved dose calculation approaches in proton therapy.

Purpose of the Study:

• To develop a deep-learning model for calculating dose-to-water (D_W) and dose-to-medium (D_M) in proton therapy.
• To achieve accuracy comparable to MC simulations with significantly reduced computation time.
• To generalize the model to different treatment sites using transfer learning.

Main Methods:

• A SwinUNetr model was trained on 259 prostate stereotactic body radiation therapy (SBRT) plans.
• Projected proton spot maps (PPSM) were generated from patient CT scans and proton spot maps (PSM).
• Transfer learning was applied to 84 central nervous system (CNS) plans for generalization.

Main Results:

• The model achieved a dose calculation time of 0.07 seconds per field on an Nvidia-A100 GPU, over 100x faster than MC.
• For prostate plans, the model yielded a mean absolute error (MAE) of 0.26 ± 0.17 Gy for D_W and 92.2% gamma passing rate.
• After transfer learning for CNS plans, MAE for D_W was 0.49 ± 0.24 Gy with an 90.1% gamma passing rate.

Conclusions:

• The SwinUNetr model offers an efficient and accurate solution for proton therapy dose distribution calculations.
• This deep learning approach has the potential to accelerate treatment planning while maintaining high accuracy.
• The model's ability to generalize across treatment sites enhances its clinical applicability.