IE-CycleGAN: improved cycle consistent adversarial network for unpaired PET image enhancement

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Image Processing

Background:

• Positron emission tomography (PET) scanners have advanced, producing higher quality images.
• State-of-the-art PET scanners are costly, limiting their availability in local hospitals.
• There is a need to enhance low-quality PET images from common scanners to match high-quality ones.

Purpose of the Study:

• To develop a method for converting low-quality PET images to high-quality images comparable to state-of-the-art scanners.
• To achieve this enhancement without requiring paired low- and high-quality PET image datasets.

Main Methods:

• Proposed an improved CycleGAN (IE-CycleGAN) model for unpaired PET image enhancement.
• Incorporated correlation coefficient loss and patient-specific prior loss to constrain generated image structure.
• Utilized a normalX-to-advanced training strategy to improve network generalization.

Main Results:

• The IE-CycleGAN method achieved results comparable to supervised methods on uEXPLORER datasets.
• On local hospital datasets, IE-CycleGAN demonstrated superior contrast-to-noise ratios (CNR) and tumor-to-background SUVmax ratios (TBR) compared to NLM, BM3D, and DIP.
• The method outperformed supervised Unet and CycleGAN on images from different scanners, indicating strong generalization.

Conclusions:

• The proposed unpaired PET image enhancement method surpasses traditional techniques like NLM, BM3D, and DIP.
• IE-CycleGAN demonstrates superior performance on local hospital datasets compared to supervised methods, highlighting its generalization capabilities.
• This approach offers a viable solution for improving PET image quality in resource-limited settings.