Pseudo PET synthesis from CT based on deep neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Integrated PET/CT imaging provides crucial anatomical and functional data for tumor diagnosis.
• High costs, limited availability, and radiation concerns associated with PET imaging restrict its clinical application.
• Developing alternative methods for functional imaging is essential to improve accessibility and reduce patient burden.

Purpose of the Study:

• To develop a cross-modal medical image synthesis method for generating PET images from CT scans.
• To accurately synthesize lesion regions in the generated PET images.
• To overcome the limitations of traditional PET imaging, such as cost and radiation exposure.

Main Methods:

• A two-stage generative adversarial network (GAN) called multi-modal fusion pre-trained autoencoder (MMF-PAE-GAN) was proposed.
• The method integrates a pre-GAN and post-GAN with a pre-trained autoencoder (PAE) for enhanced feature extraction.
• Multi-modal feature fusion and a perceptual loss were employed to improve the fidelity of synthesized lesion regions.

Main Results:

• The MMF-PAE-GAN achieved high performance on the AutoPET and FAHSU datasets, with metrics including PSNR up to 29.17 dB and SSIM up to 0.92.
• Pixel-level metrics demonstrated strong image reconstruction accuracy.
• Slice-level classification metrics showed high sensitivity (85.31%), specificity (97.02%), and accuracy (95.97%) for lesion detection.

Conclusions:

• The MMF-PAE-GAN successfully generates high-quality PET images from CT scans without radioactive tracers.
• This approach has the potential to significantly improve the accessibility and cost-effectiveness of functional imaging in clinical settings.
• The method offers a viable alternative for scenarios where PET acquisition is limited or repeated scans are not feasible.