A plan verification platform for online adaptive proton therapy using deep learning-based Monte-Carlo denoising.

Area of Science:

• Medical Physics
• Radiotherapy
• Computational Biology

Background:

• Monte Carlo (MC) simulations are crucial for accurate dose calculation in proton therapy.
• Traditional MC methods are computationally intensive, limiting their use in real-time applications like online adaptive proton therapy (APT).
• Deep learning (DL) offers potential for accelerating complex simulations.

Purpose of the Study:

• To develop a fast plan verification platform using a deep learning (DL)-based denoising approach combined with Monte Carlo (MC) methods.
• To maintain the accuracy of MC dose calculations while significantly reducing computation time.
• To investigate the platform's suitability for online adaptive proton therapy (APT).

Main Methods:

• A Monte Carlo (MC) platform for proton therapy was modeled and validated with measured data.
• A DL-based denoising model utilizing deep ResNet-deconvolution networks was developed to accelerate dose calculations.
• The DL model was trained on MC dose distributions from 52 patients, comparing low-particle (input) and high-particle (reference) simulations.

Main Results:

• The MC model demonstrated good agreement with measured data (range <0.85 mm, lateral dose profile <2.41%).
• The DL denoising approach significantly improved dose accuracy, reducing RMSE by 3.94 times and achieving a 99% gamma passing rate (3 mm/3%) compared to 82% for the un-denoised input.
• The computation time was reduced to under 60 seconds, a substantial decrease from the >100 minutes required for high-accuracy MC simulations.

Conclusions:

• An end-to-end fast plan verification platform combining MC and DL methods was successfully developed.
• The platform achieves MC-level dose calculation accuracy with significantly reduced computation time.
• This approach is suitable for online APT, offering an efficient alternative for online plan verification.