3D conditional generative adversarial networks for high-quality PET image estimation at low dose.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiochemistry

Background:

• Positron emission tomography (PET) is crucial for visualizing biochemical and physiological processes.
• High-quality PET imaging typically requires full doses of radioactive tracers, raising health concerns.
• Low-dose PET imaging increases image noise, compromising diagnostic quality.

Purpose of the Study:

• To develop a novel method for estimating high-quality, full-dose PET images from low-dose scans.
• To reduce radiation exposure for patients undergoing PET imaging.
• To maintain diagnostic image quality despite dose reduction.

Main Methods:

• A 3D conditional generative adversarial network (3D c-GAN) was developed, utilizing a U-net-like generator with skip connections.
• The generator network synthesizes full-dose PET images conditioned on low-dose inputs.
• Training incorporated estimation error loss and discriminator feedback, with a progressive refinement scheme for enhanced quality.

Main Results:

• The proposed 3D c-GAN method successfully estimated high-quality full-dose PET images from low-dose data.
• The method demonstrated superior performance compared to benchmark and state-of-the-art techniques.
• Validation on human brain datasets (normal and MCI subjects) confirmed qualitative and quantitative improvements.

Conclusions:

• The 3D c-GAN approach effectively reduces radiation dose in PET imaging while preserving essential image quality.
• This method offers a promising solution for safer and more effective PET diagnostics.
• The technique shows significant potential for clinical application in various neurological conditions.