Proton spot dose estimation based on positron activity distributions with neural network.

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Positron emission tomography (PET) is explored for reconstructing proton-induced positron activity in proton therapy.
• This technique offers potential for real-time range verification and in vivo dose monitoring.
• Deep learning for dose estimation from PET data shows promise for guided proton therapy.

Purpose of the Study:

• Evaluate recurrent neural network (RNN), U-Net, and Transformer models for proton dose estimation.
• Investigate model characteristics to guide clinical selection for proton therapy.
• Assess performance in homogeneous and heterogeneous anatomical sites.

Main Methods:

• Simulated proton spot beams using Geant4 and CT images from head cases.
• Trained neural networks using 1D positron activity distributions as input and 1D dose distributions as output.
• Examined impact of training sample size, anatomical site, and activity distribution uncertainty on dose prediction accuracy using MRE and ARE metrics.

Main Results:

• U-Net demonstrated strong range verification with few samples (75% AREs < 0.5 mm with 500 samples).
• All models performed better in homogeneous brain sites (90% AREs < 0.5 mm) than heterogeneous nasopharynx (88% AREs < 3 mm).
• Transformer achieved the best overall dose prediction (92% MREs < 3% in brain, 85% MREs < 5% in nasopharynx) and was robust to activity distribution uncertainty.

Conclusions:

• U-Net and Transformer models show distinct advantages in proton dose estimation.
• U-Net is suitable for range verification, especially with limited training data.
• Transformer excels in overall dose prediction, supporting its use in dose-guided proton therapy.