Diffusion Transformer-based Universal Dose Denoising for Pencil Beam Scanning Proton Therapy.

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence in Healthcare

Background:

• Intensity-modulated proton therapy (IMPT) provides precise tumor coverage and organ sparing in head and neck cancers.
• Anatomical changes necessitate frequent online adaptive radiation therapy (oART) for IMPT.
• Accurate and rapid dose calculation, typically via Monte Carlo (MC) simulations, is crucial for oART but is computationally intensive.

Purpose of the Study:

• To develop a diffusion transformer-based framework for denoising low-statistics MC dose maps.
• To enable fast, high-quality dose generation for online adaptive proton therapy.
• To improve the efficiency and accuracy of dose calculations in IMPT.

Main Methods:

• A diffusion transformer model was developed to denoise MC dose maps generated from IMPT plans.
• Noisy (1 min) and high-statistics (10 min) dose maps were created using MCsquare for 80 head and neck cancer patients.
• The model was trained using noisy dose maps and CT images, with high-statistics maps as ground truth, and validated on multiple cancer sites.

Main Results:

• The denoising framework achieved low Mean Absolute Errors (MAE) across various cancer sites (e.g., 0.195 Gy[RBE] for H&N, 0.120 Gy[RBE] for lung).
• High 3D Gamma passing rates (>92% at 3%/2mm) were achieved, indicating excellent agreement with high-statistics calculations.
• Dose-Volume Histogram (DVH) indices for target volumes and organs at risk closely matched the ground truth.

Conclusions:

• The developed diffusion transformer-based denoising framework effectively generates high-quality MC dose maps.
• The model demonstrates generalization capabilities, performing well across different cancer sites beyond its training data (H&N).
• This approach facilitates faster, accurate dose calculations, supporting efficient online adaptive proton therapy.