TransDose: a transformer-based UNet model for fast and accurate dose calculation for MR-LINACs.

Area of Science:

• Medical Physics
• Artificial Intelligence in Radiation Therapy
• Radiotherapy Treatment Planning

Background:

• Accurate dose calculation is critical for radiotherapy, especially with advanced technologies like magnetic resonance-linear accelerators (MR-LINACs).
• Traditional dose calculation methods can be time-consuming, limiting real-time treatment adjustments.
• The integration of artificial intelligence offers a promising avenue for accelerating dose calculations.

Purpose of the Study:

• To introduce TransDose, a novel transformer-based UNet model designed for rapid and precise dose calculation in MR-LINACs.
• To evaluate the accuracy and efficiency of TransDose compared to established methods.

Main Methods:

• Developed TransDose, a model combining a 3D residual UNet with a transformer encoder to process volumetric spatial features and long-range dependencies.
• Trained and validated the model on 98 patient cases across four tumor sites (brain, nasopharynx, lung, rectum) using Monte Carlo simulations for ground truth.
• Augmented the dataset by rotating beam angles and recalculating doses.

Main Results:

• TransDose demonstrated high accuracy, with average 3D gamma-passing rates (3%/2 mm) exceeding 98% across all tumor sites.
• The model achieved an average dose calculation time of less than 310 ms per beam, significantly faster than traditional methods.
• Dose-volume histograms and indices showed good consistency between predicted and Monte Carlo doses.

Conclusions:

• The developed TransDose model provides accurate and efficient dose calculations for MR-LINACs.
• Its speed and precision suggest significant potential for real-time applications, such as online adaptive radiotherapy plan optimization.