Generation of Conventional ¹⁸F-FDG PET Images from ¹⁸F-Florbetaben PET Images Using Generative Adversarial Network: A...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• 18F-fluorodeoxyglucose (FDG) positron emission tomography (PET) visualizes brain neuronal injury in Alzheimer's disease (AD).
• Early-phase amyloid PET imaging shows similarities to FDG-PET.
• Generating FDG-PET images from amyloid PET could offer diagnostic advantages.

Purpose of the Study:

• To generate 18F-fluorodeoxyglucose PET (PET_FDG) images from 18F-florbetaben PET (PET_FBB) images using a generative adversarial network (GAN).
• To compare the quality of generated PET_FDG (PET_GE-FDG) with real PET_FDG (PET_RE-FDG) images.
• To evaluate image quality using Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR).

Main Methods:

• Utilized the Alzheimer's Disease Neuroimaging Initiative (ADNI) database with 110 participants.
• Employed a GAN with a U-Net architecture for image-to-image translation.
• Compared cycleGAN and pix2pix models using SSIM and PSNR metrics on training and validation datasets.

Main Results:

• The cycleGAN model achieved a mean SSIM of 0.768 ± 0.135 and PSNR of 32.4 ± 9.5.
• The pix2pix model achieved a mean SSIM of 0.745 ± 0.143 and PSNR of 30.7 ± 8.0.
• cycleGAN demonstrated statistically significant higher SSIM and PSNR values compared to pix2pix (p < 0.001).

Conclusions:

• Deep learning, specifically cycleGAN, can effectively generate PET_FDG images from PET_FBB images.
• Generated PET_FDG images show high similarity to real PET_FDG images.
• This technique may aid Alzheimer's disease management by providing valuable imaging data without increased radiation or cost.