Reconstructing delivered dose in real time: a beam physics-embedded, language-model-driven approach

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning in Healthcare

Background:

• Accurate dose reconstruction is crucial for adaptive radiotherapy and intra-fraction dose monitoring.
• Current methods may lack the speed and precision required for real-time applications.

Purpose of the Study:

• To develop a fast, accurate, physics-informed framework for 3D dose reconstruction directly from machine delivery beams.
• To enable real-time dose monitoring and support adaptive radiotherapy workflows.

Main Methods:

• A two-stage, beam physics-driven dose reconstruction framework was developed.
• Stage 1 used a large language model (LLM) to extract beam characteristics as priors.
• Stage 2 employed a multimodal neural network integrating CT images, fluence, and LLM-derived priors for dose estimation.

Main Results:

• The framework achieved a mean 3D gamma passing rate of 0.975, outperforming existing deep learning models.
• Reconstruction of a 9-field plan took approximately 6.3 seconds.
• External testing demonstrated robust performance with a gamma passing rate of 0.953.

Conclusions:

• The LLM-enhanced framework improves dosimetric accuracy and maintains rapid inference speeds.
• This approach shows significant potential for adaptive radiotherapy and real-time 3D in vivo dosimetry.