Generative Adversarial Network-Enhanced Ultra-Low-Dose [¹⁸F]-PI-2620 τ PET/MRI in Aging and Neurodegenerative...

Area of Science:

• Neuroimaging
• Medical Physics
• Artificial Intelligence

Background:

• Hybrid tau PET/MR imaging is crucial for neurodegenerative disease management.
• Ultra-low-dose imaging presents challenges in image quality for clinical applications.

Purpose of the Study:

• To investigate deep learning for enhancing ultra-low-dose [18F]-PI-2620 tau PET/MR images.
• To assess if enhanced images achieve diagnostic quality comparable to full-dose scans.

Main Methods:

• A generative adversarial network was trained using ultra-low-dose tau PET/MR data (5% of full dose).
• MR images served as additional input channels for the deep learning model.
• Image quality was evaluated using region-based analyses and a reader study against full-dose images.

Main Results:

• Deep learning significantly reduced noise in ultra-low-dose tau images.
• Bias in standard uptake value ratios was reduced in enhanced images, particularly in lower uptake regions.
• Radiotracer uptake patterns in enhanced images were accurately interpreted compared to full-dose scans.

Conclusions:

• Deep learning-enhanced ultra-low-dose tau PET/MR imaging yields diagnostic quality comparable to full-dose imaging.
• This approach may allow for reduced radiation exposure and more frequent dementia monitoring.
• Clinical readings were consistent, supporting the potential for dose reduction in tau PET/MR.