Whole-body CT-to-PET synthesis using a customized transformer-enhanced GAN

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• 18F-FDG PET-CT is a valuable diagnostic tool, but faces limitations including long scan times, high costs, and significant radiation exposure.
• Computed tomography (CT) provides anatomical detail, while positron emission tomography (PET) reveals metabolic activity, making their integration powerful for lesion and tumor detection.

Purpose of the Study:

• To develop a deep learning model for whole-body CT-to-PET synthesis, aiming to generate high-quality synthetic PET images.
• To create synthetic PET images that are clinically relevant and diagnostically equivalent to real PET scans.

Main Methods:

• A transformer-enhanced generative adversarial network (CPGAN) was developed for synthesizing PET images from CT scans.
• The model incorporates residual blocks and fully connected transformer residual blocks to capture both local and global contextual information.
• A custom loss function emphasizing structural consistency was implemented to enhance the quality of synthesized PET images.

Main Results:

• The CPGAN model demonstrated superior performance compared to seven other state-of-the-art models in quantitative evaluations (NRMSE, PSNR, SSIM).
• Quantitative metrics on the test set showed mean NRMSE of (16.90±12.27)×10-4, PSNR of 28.71±2.67, and SSIM of 0.926±0.033.
• Blind evaluation by three radiologists found no statistical difference between 50 real and 50 synthetic PET images, indicating comparable diagnostic value.

Conclusions:

• The CPGAN model successfully synthesizes high-quality PET images from CT scans, overcoming the limitations of CT in reflecting metabolic information.
• This AI-driven approach holds significant potential for reducing the need for actual PET-CT scans in clinical practice.
• The synthesized PET images maintain clinical relevance and diagnostic value, offering a promising alternative for medical imaging.