Synthesis of Amyloid Images Using a Generative Adversarial Network from 2-Dimensional ¹⁸F-FDG Images and Evaluation...

Area of Science:

• Nuclear medicine
• Artificial intelligence in medical imaging
• Alzheimer disease diagnostics

Background:

• Amyloid PET scans are increasingly used to identify Alzheimer disease patients eligible for disease-modifying therapies.
• Reducing the number of amyloid PET scans could lower healthcare costs and patient burden.

Purpose of the Study:

• To develop and evaluate a generative AI algorithm for synthesizing amyloid PET images from ¹⁸F-FDG PET images.
• To assess the feasibility of using AI-generated amyloid PET images to reduce the need for actual amyloid PET scans.

Main Methods:

• A 2-dimensional pix2pix generative adversarial network algorithm was employed.
• The algorithm was trained and validated on paired ¹⁸F-FDG PET and amyloid PET images from 55 patients.
• Evaluation metrics included image quality, voxel values, white/gray matter contrast, and diagnostic classification performance.

Main Results:

• Synthetic amyloid PET images were visually consistent with real scans, preserving anatomical continuity.
• Voxel values and white/gray matter contrast in synthetic images showed strong correlations with real images.
• A 2-class classifier achieved over 85% accuracy in detecting β-amyloid (Aβ) deposition using synthetic images.

Conclusions:

• Generative AI can successfully synthesize amyloid PET images from ¹⁸F-FDG PET data, capturing key diagnostic features.
• The synthetic images demonstrated acceptable accuracy (85%) for classifying Aβ deposition.
• This AI-driven approach holds potential for clinical application to reduce unnecessary amyloid PET scans.