UNetrDose: a fast and accurate transformer-based dose prediction model for radiotherapy

Area of Science:

• Medical Physics
• Artificial Intelligence in Radiation Oncology
• Deep Learning for Medical Imaging

Background:

• Accurate dose calculation is crucial for intensity-modulated radiation therapy (IMRT).
• Traditional methods can be computationally intensive, limiting efficiency in clinical workflows.
• Transformer-based deep learning offers potential for rapid and precise dose prediction.

Purpose of the Study:

• To develop and validate UNetrDose, a Transformer-based deep learning model for fast and accurate photon beamlet dose prediction.
• To achieve Monte Carlo (MC)-level dosimetric accuracy using simplified inputs (CT images, beamlet coordinates).
• To enable efficient reconstruction of 3D dose distributions for IMRT plans.

Main Methods:

• A Transformer-based UNetrDose model was developed, integrating convolutional layers and Transformer modules.
• Input data included 3D CT patches and beamlet geometric information.
• Ground-truth doses were generated using MC simulations for 90 IMRT plans (esophageal and rectal cases).
• Performance was evaluated using 3D gamma pass rates and Dose-Volume Histogram (DVH) comparisons.

Main Results:

• UNetrDose achieved >96% pass rates for the stringent γ(1 mm, 1%) criterion at the beamlet level.
• Mean γ(2 mm, 2%) pass rates for full IMRT plans were high, ranging from 97.06% to 98.75%.
• The model demonstrated high efficiency with an average inference time of ~28 ms per beamlet.

Conclusions:

• UNetrDose provides a balance of high dosimetric accuracy and computational speed, serving as a viable alternative to traditional dose calculation engines.
• The model's simplified input requirements and fast prediction capabilities enhance clinical workflow efficiency for IMRT.
• This deep learning approach shows significant promise for time-sensitive radiation therapy planning scenarios.