Dose calculation in proton therapy using a discovery cross-domain generative adversarial network (DiscoGAN).

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning Applications

Background:

• Accurate dose calculation is fundamental for effective proton therapy.
• Monte Carlo simulations provide high accuracy but are computationally intensive.
• There is a need for faster, accurate dose calculation methods in proton therapy.

Purpose of the Study:

• To develop a novel machine learning-based approach for proton therapy dose calculation.
• To achieve accuracy comparable to Monte Carlo simulations while reducing computational time.
• To evaluate the efficacy of a Discovery Cross-Domain Generative Adversarial Network (DiscoGAN) for this purpose.

Main Methods:

• Utilized computed tomography (CT)-based patient phantoms for thorax, head, and abdomen treatment sites.
• Generated training data using Monte Carlo simulations.
• Developed a DiscoGAN to map CT image HU values to dose distributions, incorporating stopping power and HU values as auxiliary features.

Main Results:

• The DiscoGAN model demonstrated effective dose calculation across different treatment sites.
• Achieved mean relative errors (MRE) as low as 1.47% (abdomen), 2.43% (thorax), and 2.83% (head).
• Independent validation showed comparable accuracy, with mean MREs around 1.64%-4.02%, and no significant dependency on beam energy (80-130 MeV).

Conclusions:

• The DiscoGAN framework shows significant potential for accurate and efficient dose calculation in proton therapy.
• The approach offers comparable accuracy to Monte Carlo simulations with reduced computational cost.
• Future research will compare DiscoGAN with the pencil beam algorithm, with potential applications in inverse planning and adaptive proton therapy.