Personalized MR-Informed Diffusion Models for 3D PET Image Reconstruction.

Area of Science:

• Medical Imaging
• Radiochemistry
• Artificial Intelligence

Background:

• Diffusion models enhance PET image reconstruction, improving lesion detectability and flexibility.
• Current methods train on noisy PET images without sinogram data.

Purpose of the Study:

• To propose a method for generating subject-specific pseudo-PET images using multi-subject PET-MR scans.
• To enhance personalized diffusion model pre-training for improved PET reconstruction accuracy with low-count data.

Main Methods:

• Synthesize pseudo-PET images by transforming anatomy between patients using image registration.
• Utilize subject's MR scan information for higher resolution and anatomical feature retention.
• Pre-train a personalized diffusion model with subject-specific pseudo-PET images.

Main Results:

• Improved reconstruction accuracy with low-count PET data using personalized diffusion models.
• Demonstrated effective combination of MR guidance without imposing excessive anatomical features.
• Achieved an improved balance between reconstructing PET-unique features and shared PET-MR features.

Conclusions:

• Generating and utilizing synthetic pseudo-PET data enhances medical imaging tasks.
• Patient-specific PET image generation is possible without deep learning or extensive datasets.
• This method offers a promising approach for improved PET-MR image reconstruction trade-offs.