Ultra-low-dose PET reconstruction using generative adversarial network with feature matching and task-specific...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuroscience

Background:

• Amyloid Positron Emission Tomography (PET) imaging is crucial for diagnosing Alzheimer's disease.
• Standard-dose PET scans require significant radiotracer injection, posing potential risks.
• Developing methods to obtain high-quality PET images from low-dose scans is essential for patient safety and diagnostic accuracy.

Purpose of the Study:

• To synthesize standard-dose amyloid PET images from ultra-low-dose PET data using a generative adversarial network (GAN).
• To ensure the synthesized images possess high quality and accurately retain pathological features indicative of amyloid.
• To evaluate the effectiveness of feature matching and task-specific perceptual loss in GAN-based image synthesis.

Main Methods:

• Utilized a 2D encoder-decoder GAN architecture for image synthesis from ultra-low-dose PET data.
• Employed multi-slice inputs for enhanced noise reduction and 2.5D information integration.
• Incorporated feature matching to minimize hallucinated structures and task-specific perceptual loss to preserve pathological features.
• Evaluated image quality using PSNR, SSIM, RMSE, and expert radiologist assessments.

Main Results:

• The proposed GAN method significantly outperformed a previous state-of-the-art PET-only model in PSNR, SSIM, and RMSE metrics.
• Achieved comparable results to a PET-MR model, demonstrating the efficacy of the PET-only approach.
• Expert radiologists confirmed superior image quality and better preservation of amyloid-related pathological features compared to existing methods.

Conclusions:

• Standard-dose amyloid PET images can be successfully synthesized from ultra-low-dose inputs using GANs.
• Adversarial learning, feature matching, and task-specific perceptual loss are critical components for ensuring image quality and pathological feature preservation.
• This technique holds promise for improving the safety and efficiency of amyloid PET imaging.