Convolution neural network toward Monte Carlo photon dose calculation in radiation therapy

Area of Science:

• Medical Physics
• Radiotherapy
• Computational Biology

Background:

• Monte Carlo (MC) simulations are highly accurate for radiotherapy dose calculations but are computationally intensive.
• The long calculation times of MC hinder its routine clinical application.
• Accurate and efficient dose calculation is crucial for optimizing radiotherapy treatment plans.

Purpose of the Study:

• To develop a fast and accurate method for calculating 3D dose distributions in radiotherapy using a convolutional neural network (CNN).
• To investigate the feasibility of using total energy release per unit mass (TERMA) and electron density (ED) distributions as inputs for the CNN.
• To reduce the computational time of dose calculations while maintaining accuracy comparable to full MC.

Main Methods:

• A CNN model (T-MC Net) was trained using MC-generated datasets of head and neck CT images with 6-MV photon beams.
• A novel Dose-mixup data augmentation technique and training strategy were employed.
• The model was tested on rectangular and intensity-modulated radiation therapy (IMRT) fields, evaluating gamma pass rates.

Main Results:

• High gamma pass rates were achieved across various criteria (1%/2 mm, 2%/2 mm, 3%/2 mm) for both rectangular and IMRT fields.
• For rectangular fields, gamma pass rates ranged from 90.11% to 99.16%.
• For IMRT fields, gamma pass rates were consistently high, reaching up to 99.63%.

Conclusions:

• The CNN-based approach successfully calculated dose distributions with accuracy comparable to full MC simulations.
• The proposed method offers a significant reduction in calculation time, making it a potential clinical tool.
• This CNN model can serve as an acceleration engine for dose algorithms, particularly beneficial for time-sensitive radiotherapy applications.