Fast spot-scanning proton dose calculation method with uncertainty quantification using a three-dimensional...

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Proton therapy offers precise dose delivery but requires accurate dose calculation.
• Real-time dose calculation and uncertainty estimation are crucial for advanced applications like image-guided proton therapy.
• Current methods can be computationally intensive, limiting near real-time applications.

Purpose of the Study:

• To develop a near-real-time spot-scanning proton dose calculation method using a 3D convolutional neural network (3D-CNN).
• To incorporate probabilistic uncertainty estimation into the dose calculation process.
• To evaluate the accuracy, efficiency, and adaptability of the 3D-CNN model for proton therapy dose prediction.

Main Methods:

• A 3D-CNN model was trained and tested on CT images and clinical target volumes from 215 head and neck cancer patients.
• Spot beam data was converted into a 'peak map' (PM) input for the 3D-CNN.
• Transfer learning was employed to adapt the model to different beam parameters and calculation algorithms with minimal data.

Main Results:

• The 3D-CNN model achieved accurate 3D proton dose distribution calculations with a mean absolute error of 0.778 cGyE.
• Predicted uncertainties correlated with dose errors at high contrast edges.
• Inference time was approximately 0.8 seconds per plan on a consumer-grade GPU, demonstrating near real-time capability.

Conclusions:

• The developed 3D-CNN method provides accurate and efficient near-real-time proton dose calculation with uncertainty estimation.
• The model is adaptable to various beam configurations and calculation settings via transfer learning.
• This approach holds significant potential for dose verification, image-guided proton therapy, and other clinical applications.