DeepDose: Towards a fast dose calculation engine for radiation therapy using deep learning.

Area of Science:

• Medical Physics
• Artificial Intelligence
• Radiation Oncology

Background:

• Accurate dose calculation is crucial for effective radiation therapy.
• Current methods can be computationally intensive, limiting real-time adaptation.
• Deep learning offers potential for faster and efficient dose prediction.

Purpose of the Study:

• To introduce DeepDose, a deep learning framework for fast and accurate dose calculations in intensity-modulated radiation therapy (IMRT).
• To validate the accuracy and speed of DeepDose against Monte Carlo simulations for prostate cancer patients.

Main Methods:

• Developed a novel framework (DeepDose) using physics-based inputs to train a deep convolutional network.
• Trained the network on dose distributions from 80 prostate cancer patients, validated on 10, and tested on 11 independent cases.
• Utilized Monte Carlo simulations for ground-truth dose calculations at 1% uncertainty and 3 mm³ grid spacing.

Main Results:

• DeepDose accurately estimated segment doses for test patients, achieving 99.9%±0.3% agreement with Monte Carlo calculations (3%/3 mm gamma index).
• Dose prediction was highly efficient, with total patient calculation time of approximately 1 minute per patient.
• The framework demonstrated proof-of-concept on 101 prostate cancer patients undergoing fixed-beam IMRT.

Conclusions:

• DeepDose provides a fast and accurate method for radiation therapy dose calculation.
• The framework's speed and accuracy are compelling for online adaptive radiation therapy workflows.
• Deep learning shows significant promise for advancing radiation treatment planning and delivery.