Robust whole-body PET image denoising using 3D diffusion models: evaluation across various scanners, tracers, and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiochemistry

Background:

• Whole-body Positron Emission Tomography (PET) imaging is crucial for cancer diagnosis and treatment.
• Current deep learning denoising methods struggle with diverse PET acquisition protocols, limiting their clinical applicability.
• Low image quality in PET hinders accurate diagnosis and effective treatment monitoring.

Purpose of the Study:

• To propose and validate a 3D Denoising Diffusion Probabilistic Model (3D DDPM) for robust and universal whole-body PET image denoising.
• To address the limitations of traditional deep learning methods in handling varied PET imaging protocols.
• To develop a foundational model for enhancing PET image quality across different clinical settings.

Main Methods:

• A 3D DDPM was developed, learning to reconstruct clean PET images from noisy inputs through a forward and reverse diffusion process.
• A 3D convolutional network was trained on high-quality PET/CT data to capture accurate distribution information.
• The model's performance was rigorously evaluated using data from four scanners, four tracer types, and six dose levels.

Main Results:

• The 3D DDPM demonstrated superior denoising performance compared to 2D DDPM, 3D UNet, and 3D GAN across all tested conditions.
• The model achieved consistently better results in various clinical scenarios, including different scanners, tracers, and dose levels.
• Uncertainty maps generated by the 3D DDPM showed lower variance, indicating higher confidence in the denoised image outputs.

Conclusions:

• The 3D DDPM is an effective and versatile solution for whole-body PET image denoising, adaptable to diverse clinical settings.
• This model establishes a promising foundation for improving PET image quality and can be readily adopted by researchers.
• The developed 3D DDPM offers an off-the-shelf solution for whole-body PET image denoising, with code and model publicly available.