Unpaired Low-Dose CT Denoising Network Based on Cycle-Consistent Generative Adversarial Network with Prior Image...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• X-ray computed tomography (CT) is crucial for diagnosis, but concerns exist about patient radiation dose.
• Reducing radiation dose in CT (low-dose CT or LDCT) introduces noise, potentially compromising diagnostic accuracy.
• Current deep learning methods for LDCT denoising often require paired datasets, which are difficult and costly to obtain.

Purpose of the Study:

• To develop an effective low-dose CT image denoising method that does not require paired training data.
• To improve the quality of LDCT images for better radiologists' judgment and confidence.
• To address the limitations of existing deep learning approaches in LDCT image reconstruction.

Main Methods:

• Proposed an unpaired LDCT image denoising network utilizing CycleGAN architecture.
• Incorporated cyclic loss for image-to-image translation between LDCT and normal-dose CT (NDCT) distributions.
• Integrated prior image information to supervise content generation and ensure accurate image detail preservation.

Main Results:

• The proposed method significantly reduces image noise in LDCT scans.
• Quantitative evaluation showed improvements in peak signal-to-noise ratio (PSNR) by over 3 dB and increased structural similarity (SSIM) compared to standard CycleGAN.
• Real-data experiments confirmed the method's superiority in both visual quality and quantitative metrics.

Conclusions:

• The developed unpaired LDCT denoising network effectively reduces noise while retaining critical image information.
• This approach overcomes the need for paired datasets, offering a practical solution for improving LDCT image quality.
• The method enhances diagnostic confidence by providing clearer, more reliable LDCT images.