Boron concentration prediction from Compton camera image for boron neutron capture therapy based on generative...

Area of Science:

• Medical Physics
• Artificial Intelligence in Medicine
• Radiotherapy Imaging

Background:

• Accurate boron concentration prediction is crucial for Boron Neutron Capture Therapy (BNCT) dose calculations.
• Current online monitoring methods for boron concentration can be time-consuming.

Purpose of the Study:

• To develop a fast and accurate method for predicting boron distribution using generative adversarial networks (GANs).
• To provide a computational basis for improving online boron monitoring in BNCT.

Main Methods:

• Simulated BNCT and Compton imaging processes.
• Trained a GAN using reconstructed Compton camera images and CT skin contours as input, with PET-derived boron concentration as output.
• Evaluated image quality using structural similarity, peak signal-to-noise ratio, and root mean square error.

Main Results:

• Significantly improved image quality metrics compared to original images.
• Enhanced the tumor-to-normal tissue boron concentration ratio from 1.55 to 3.85 (closer to the true value of 3.52).
• Achieved image optimization in under 0.83 seconds, outperforming iterative methods.

Conclusions:

• The proposed GAN-based method offers a rapid computational solution for online boron concentration monitoring in BNCT.
• This approach can accelerate the clinical application and development of BNCT technology.