Neural network-driven direct CBCT-based dose calculation for head-and-neck proton treatment planning.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Accurate proton dose calculation is crucial for adaptive radiotherapy, especially with anatomical changes detected via cone beam computed tomography (CBCT).
• Traditional CBCT-based dose calculations face limitations due to image quality, necessitating complex correction workflows.

Purpose of the Study:

• To develop and validate a deep learning approach for direct proton dose calculation from CBCT images.
• To utilize extended long short-term memory (xLSTM) neural networks to overcome traditional CBCT dose calculation challenges.

Main Methods:

• A retrospective dataset of 40 head-and-neck cancer patients was used to train an xLSTM-based neural network (CBCT-NN).
• The network incorporated energy token encoding and beam's-eye-view sequence modeling.
• Training involved 82,500 proton pencil beam configurations with Monte Carlo (MC)-generated ground truth doses; validation used gamma analysis, MPDE, and DVH comparison.

Main Results:

• The CBCT-NN achieved high accuracy with a 95.1 ± 2.7% gamma pass rate (2mm/2% criteria).
• Mean percentage dose errors (MPDE) were 2.6 ± 1.4% (high-dose regions) and 5.9 ± 1.9% (global).
• Excellent preservation of target coverage and organ-at-risk constraints was observed, with computation time under 3 minutes.

Conclusions:

• This study demonstrates the feasibility of direct CBCT-based proton dose calculation using xLSTM neural networks.
• The developed approach offers comparable accuracy and computational efficiency to MC methods, suitable for adaptive radiotherapy protocols.
• This deep learning method eliminates the need for traditional CBCT correction workflows.