Study of low-dose PET image recovery using supervised learning with CycleGAN.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Positron Emission Tomography (PET) is crucial for clinical diagnosis and radiation therapy.
• Low-dose PET (LDPET) minimizes radiation exposure but can compromise image quality.
• Developing methods to enhance LDPET image quality is essential for safe and effective clinical use.

Purpose of the Study:

• To propose a novel supervised deep learning model, S-CycleGAN, for recovering low-dose PET brain images.
• To evaluate the performance of S-CycleGAN against other deep learning methods (RED-CNN, 3D-cGAN).
• To assess the accuracy, efficiency, and robustness of S-CycleGAN in improving LDPET image quality and diagnostic metrics.

Main Methods:

• A supervised deep learning approach using a Generative Adversarial Network (GAN) with cycle-consistency loss, Wasserstein distance loss, and supervised learning loss (S-CycleGAN).
• Application of S-CycleGAN, RED-CNN, and 3D-cGAN to 10% and 30% dose testing datasets and simulated datasets with varying lesion characteristics.
• Quantitative evaluation using NRMSE, SSIM, PSNR, LPIPS, SUVmax, and SUVmean, alongside qualitative visual comparisons.

Main Results:

• S-CycleGAN demonstrated comparable SSIM and PSNR to other methods.
• The proposed model showed slightly higher noise but improved perception scores and preserved image details.
• S-CycleGAN significantly outperformed RED-CNN and 3D-cGAN in SUVmean and SUVmax metrics.

Conclusions:

• The S-CycleGAN approach is accurate, efficient, and robust for recovering low-dose PET brain images.
• This deep learning method offers a promising solution for enhancing LDPET image quality and diagnostic utility.
• S-CycleGAN effectively balances image quality improvement with reduced radiation dosage in PET imaging.