Multistage Diffusion Model With Phase Error Correction for Fast PET Imaging.

Area of Science:

• Medical Imaging
• Deep Learning
• Radiochemistry

Background:

• Fast Positron Emission Tomography (PET) imaging is crucial for minimizing motion artifacts and enhancing patient comfort.
• Current diffusion-based deep learning methods struggle with accurate PET degradation modeling, error accumulation, and require numerous reconstruction iterations.

Purpose of the Study:

• To develop a novel multistage diffusion framework for high-fidelity fast PET imaging.
• To address limitations of existing methods in capturing PET degradation and reducing reconstruction time.

Main Methods:

• A multistage diffusion framework was designed to approximate non-linear temporal PET degradation using paired PET images.
• A Phase Error Correction Network (PECNet) was implemented to maintain consistency across diffusion stages.
• A deterministic cold diffusion mechanism was introduced for intra-stage degradation simulation, reducing iterations to 10.

Main Results:

• The framework achieved peak PSNRs of 36.2 dB for [68Ga]FAPI and 39.0 dB for [18F]FDG PET.
• Average SSIM values exceeded 0.97 across datasets.
• The method demonstrated superior performance compared to existing approaches in evaluations.

Conclusions:

• The proposed multistage diffusion framework enables high-fidelity PET imaging with significantly fewer reconstruction iterations.
• This approach offers a practical solution for accelerated clinical PET imaging, improving efficiency and potentially patient outcomes.