Enhancing image quality in fast neutron-based range verification of proton therapy using a deep learning-based prior...

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Computational Imaging

Background:

• Proton therapy offers precise dose delivery but requires accurate range verification.
• Fast neutron (FN) detection is a promising method for real-time range verification.
• Image reconstruction challenges, including noise and detector effects, impact accuracy.

Purpose of the Study:

• To investigate the efficacy of list-mode maximum a posteriori (LM-MAP) expectation maximization (EM) with deep learning-based priors for FN-based proton therapy range verification.
• To evaluate the performance of a conditional generative adversarial network (pix2pix) in predicting FN production distributions.
• To compare different LM-MAP-EM reconstruction strategies under varying noise and detector resolution conditions.

Main Methods:

• A conditional generative adversarial network (pix2pix) was trained to estimate FN production distributions from back-projection images.
• Monte Carlo simulations generated FN data for an 85 MeV proton beam in a lung cancer patient model, simulating range shifts.
• LM-MAP-EM algorithms were implemented, incorporating a pix2pix-predicted mean prior, and compared against LM-MLEM and conventional LM-MAP-EM.

Main Results:

• The pix2pix network showed degradation in performance with introduced detector resolution effects and noise.
• LM-MAP-EM incorporating a mean prior estimate significantly improved reconstruction performance compared to LM-MLEM.
• The LM-MAP-EM with a mean prior estimate outperformed conventional LM-MAP-EM with a quadratic energy function in SSIM.

Conclusions:

• Deep learning-based prior information can enhance iterative image reconstruction for proton therapy range verification.
• The robustness of deep learning models is constrained by data quality, particularly in high-noise scenarios.
• Further research is needed to improve deep learning model performance under realistic, noisy conditions in proton therapy applications.