UR-cycleGAN: Denoising full-body low-dose PET images using cycle-consistent Generative Adversarial Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiochemistry

Background:

• Low-dose PET (LDPET) imaging is crucial for reducing radiation exposure and scan time.
• However, LDPET images often suffer from reduced image quality and increased noise.
• Enhancing LDPET image quality is essential for accurate diagnosis and patient management.

Purpose of the Study:

• To develop and evaluate a CycleGAN-based denoising model for LDPET image enhancement.
• To achieve image quality comparable to standard-dose PET (SDPET) images.
• To significantly reduce PET imaging acquisition time.

Main Methods:

• A UR-CycleGAN model was developed using fluorine-18 fluorodeoxyglucose (18F-FDG) PET/CT data from 37 patients.
• Low-dose images were simulated with a 30-second acquisition time, while standard-dose images used a 2.5-minute acquisition.
• The model was trained on 13,210 image pairs and evaluated using PSNR and SSIM metrics.

Main Results:

• The CycleGAN-based denoising model significantly improved the quality of simulated LDPET images.
• Denoised images showed a clear trend towards the quality of SDPET images.
• Objective evaluation metrics (PSNR, SSIM) confirmed the enhancement in image quality.

Conclusions:

• The proposed method achieves high-quality PET imaging with an 80% reduction in acquisition time.
• The model enhances visual detail fidelity, demonstrating practical utility.
• This approach offers a feasible solution for reducing imaging time while maintaining diagnostic image quality.