Sub-second photon dose prediction via transformer neural networks.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Fast dose calculation is crucial for adaptive radiotherapy workflows.
• Current physics-based algorithms trade accuracy for speed.
• Deep learning offers potential for high-fidelity and rapid dose prediction.

Purpose of the Study:

• To present a deep learning algorithm, iDoTA, for accurate photon beam dose prediction.
• To achieve millisecond-level dose calculation speeds by combining transformer and convolutional layers.

Main Methods:

• The improved Dose Transformer Algorithm (iDoTA) uses a sequence modeling approach for 3D dose prediction.
• It combines a Transformer backbone for long-range dependencies and 3D convolutions for local features.
• Trained on 1700 beam dose distributions from clinical VMAT plans for diverse cancer types.

Main Results:

• iDoTA predicts individual photon beams in approximately 50 ms with a 97.72% (2 mm, 2%) gamma pass rate.
• Full VMAT dose distributions are estimated in 6-12 seconds with a 99.51% (2 mm, 2%) gamma pass rate.
• Achieved state-of-the-art performance with an average relative dose error of 0.75 ± 0.36%.

Conclusions:

• iDoTA achieves millisecond-speed prediction per beam, essential for online adaptive treatments.
• Represents a new state-of-the-art in data-driven photon dose calculation.
• Significantly accelerates photon dose calculation workflows from minutes to seconds.