PET Image Reconstruction Using Deep Diffusion Image Prior.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Diffusion models show potential in medical imaging but face challenges in Positron Emission Tomography (PET) due to tracer variability and computational costs.
• Existing PET reconstruction methods struggle with tracer-specific contrast and high computational demands, limiting their widespread application.

Purpose of the Study:

• To develop an anatomical prior-guided PET image reconstruction method using diffusion models to address limitations in tracer-specific contrast and computational efficiency.
• To enable high-quality PET image reconstruction from various tracers using a single, pretrained score function.

Main Methods:

• Proposed an anatomical prior-guided PET reconstruction method based on the deep diffusion image prior (DDIP) framework.
• Employed alternating diffusion sampling and model fine-tuning guided by PET sinogram data.
• Integrated the half-quadratic splitting (HQS) algorithm for improved computational efficiency by decoupling network optimization from iterative reconstruction.

Main Results:

• Demonstrated robust generalization across different PET tracers and scanner types using models pretrained on other tracers.
• Achieved high-quality image reconstruction from low-dose PET data, including out-of-distribution (OOD) performance assessment.
• Validated the method on simulation and clinical datasets ([18F]FDG and [18F]Florbetapir).

Conclusions:

• The proposed method offers an efficient and versatile framework for low-dose PET image reconstruction.
• The anatomical prior-guided diffusion model successfully reconstructs high-quality PET images across various tracers and conditions.
• This approach enhances the applicability of diffusion models in PET imaging, overcoming previous limitations.