DIFFUSION MODEL-BASED POSTERIOR DISTRIBUTION PREDICTION FOR KINETIC PARAMETER ESTIMATION IN DYNAMIC PET.

Area of Science:

• Neuroimaging
• Molecular Imaging
• Artificial Intelligence in Medicine

Background:

• Positron Emission Tomography (PET) enables molecular imaging of processes like hyperphosphorylated tau (p-tau) in neurodegenerative diseases.
• Tracer kinetic modeling quantifies p-tau density and cerebral perfusion from PET data, but image noise introduces uncertainty.
• Bayesian inference and Markov Chain Monte Carlo (MCMC) methods estimate parameter uncertainty but are computationally intensive.

Purpose of the Study:

• To develop a computationally efficient deep learning method for inferring posterior distributions of kinetic parameters from PET data.
• To leverage denoising diffusion probabilistic models (DDPM) for improved uncertainty quantification in PET imaging analysis.
• To evaluate the performance of the DDPM approach against traditional MCMC methods in a [18F]MK6240 PET study.

Main Methods:

• Implementation of a novel denoising diffusion probabilistic model (DDPM) for PET image analysis.
• Application of the DDPM to estimate kinetic parameters and their posterior distributions from [18F]MK6240 PET data.
• Comparison of the DDPM method's computational time, accuracy, and precision against a standard MCMC approach.

Main Results:

• The DDPM approach achieved over 30 times faster computation compared to MCMC.
• The method demonstrated high accuracy, with a mean error consistently below 0.8%.
• Posterior distributions were predicted with high precision, showing a standard deviation error below 5.77%.

Conclusions:

• Deep learning, specifically DDPM, offers a significant computational advantage for PET data analysis.
• The proposed DDPM method provides accurate and precise quantification of kinetic parameters, crucial for neurodegenerative disease research.
• This approach enhances the efficiency of molecular imaging analysis, potentially accelerating research into diseases like Alzheimer's.