A deep learning based dynamic arc radiotherapy photon dose engine trained on Monte Carlo dose distributions

Area of Science:

• Medical Physics
• Artificial Intelligence in Radiation Therapy
• Computational Dosimetry

Background:

• Monte Carlo (MC) dose engines are crucial for accurate radiotherapy planning but demand extensive computation time.
• Hardware acceleration has not fully overcome the speed limitations of traditional MC methods for reducing stochastic noise.

Purpose of the Study:

• To develop and validate a deep learning (DL) based dose engine for rapid and accurate dose distribution calculation in radiotherapy.
• To significantly reduce computation time compared to conventional MC methods.

Main Methods:

• A neural network combining 2D convolution and recurrence was developed and trained on 350 radiotherapy treatment plans.
• Dose distributions were computed using MC for 6 MV and 10 MV beams, including dynamic arcs with motion.
• Model parameters were optimized to minimize the difference between MC and DL computed doses.

Main Results:

• The DL dose engine achieved an average speed increase of 82 times over MC computation at 1% accuracy.
• Global gamma pass rates (2%/2mm) were 99.6% for doses >10% max, with mean local gamma within 2%.
• Accuracy in the high dose region (>50% max) approached 1%.

Conclusions:

• A DL-based dose engine can accurately replicate MC-computed dynamic arc radiotherapy dose distributions.
• The developed DL engine offers a high-speed solution for radiotherapy dose calculation, improving efficiency without compromising accuracy.