Frequency-Domain-Based Structure Losses for CycleGAN-Based Cone-Beam Computed Tomography Translation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• CycleGAN enables synthetic medical image generation from unpaired data.
• Artifacts in CycleGAN-generated images limit their clinical reliability.
• Improving image quality for cone-beam computed tomography (CBCT) to computed tomography (CT) translation is crucial.

Purpose of the Study:

• To address CycleGAN artifacts in medical imaging.
• To propose and evaluate a generalized frequency-based loss for CycleGAN.
• To enhance the translation of CBCT to CT-like quality synthetic CT (sCT) images.

Main Methods:

• Explored the impact of structure losses on CycleGAN.
• Developed a generalized frequency-based loss to preserve frequency domain content.
• Applied the proposed loss to CBCT to CT image translation.
• Compared generated sCT images against baseline CycleGAN and other structure losses.

Main Results:

• Proposed methods quantitatively and qualitatively improved over baseline CycleGAN across all metrics (MAE, MSE, NMSE, PSNR, SSIM).
• Achieved superior performance compared to existing structure losses.
• Generated sCT images exhibited no observable artifacts or loss in image quality.
• Demonstrated superior performance of generated sCTs over original CBCT on downstream tasks.

Conclusions:

• The generalized frequency-based loss effectively reduces artifacts in CycleGAN-generated medical images.
• This novel approach enhances the quality and reliability of synthetic CT images.
• The improved sCT images show potential for better performance in downstream clinical applications.