PET image denoising based on denoising diffusion probabilistic model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Positron Emission Tomography (PET) image quality is often compromised by physical degradation and low photon counts.
• Existing denoising methods struggle to fully restore PET image fidelity.

Purpose of the Study:

• To propose and evaluate denoising diffusion probabilistic model (DDPM)-based methods for enhancing PET image quality.
• To investigate the impact of incorporating prior imaging information within the DDPM framework.

Main Methods:

• Developed and tested DDPM frameworks for PET image denoising using [18F]FDG and [18F]MK-6240 brain datasets.
• Explored strategies including direct PET image input and using prior images (e.g., MRI) as network input or constraints.

Main Results:

• DDPM-based methods significantly outperformed nonlocal mean, Unet, and Generative Adversarial Network (GAN) denoising techniques.
• Integrating Magnetic Resonance (MR) prior information improved performance and reduced uncertainty.
• The optimal approach involved using MR prior as network input with PET data as a consistency constraint during inference.

Conclusions:

• DDPM offers a flexible and effective framework for PET image denoising.
• DDPM-based approaches surpass traditional and GAN-based methods, particularly when leveraging prior imaging data.