Diffused Multi-scale Generative Adversarial Network for low-dose PET images reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Low-dose PET (L-PET) imaging reduces radiation exposure but often compromises image quality and diagnostic performance.
• Developing methods to reconstruct high-quality full-dose PET (F-PET) images from L-PET data is crucial for clinical applications.

Purpose of the Study:

• To develop and validate a novel generative adversarial network, the Diffused Multi-scale Generative Adversarial Network (DMGAN), for converting L-PET to F-PET images.
• To achieve a balance between reduced radiation dose and preserved diagnostic accuracy in PET imaging.

Main Methods:

• The proposed DMGAN model comprises a diffusion generator and a u-net discriminator.
• The diffusion generator extracts multi-scale information to enhance generalization and training stability.
• The u-net discriminator refines generated images by analyzing details from both global and local perspectives.

Main Results:

• The DMGAN method achieved superior performance, outperforming other methods in quantitative evaluations using Structure Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR).
• The synthesized F-PET images demonstrated at least a 6.2% improvement in PSNR compared to alternative methods.
• The DMGAN-generated images exhibited more accurate voxel-wise metabolic intensity distribution, leading to clearer visualization of epilepsy foci.

Conclusions:

• The DMGAN model effectively restores image details from L-PET data, outperforming existing models.
• This approach offers a promising solution for minimizing radiation exposure in PET scans without sacrificing diagnostic utility.