Generative Consistency Models for Estimation of Kinetic Parametric Image Posteriors in Total-Body PET

Area of Science:

• Medical Imaging
• Computational Biology
• Radiochemistry

Background:

• Dynamic total body positron emission tomography (TB-PET) allows simultaneous kinetic tracer measurement across the entire body.
• Whole-body kinetics are complex and require accurate parametric imaging with statistical uncertainty.
• Current Bayesian methods like Markov chain Monte Carlo (MCMC) are computationally too intensive for TB-PET.

Purpose of the Study:

• To introduce a novel generative consistency model (CM) for efficient posterior sampling of kinetic parameters in TB-PET.
• To overcome the computational limitations of traditional Bayesian techniques for whole-body kinetic analysis.
• To enable routine, fully Bayesian parametric imaging in TB-PET applications.

Main Methods:

• Developed a generative consistency model (CM) for posterior distribution sampling of kinetic parameters.
• Evaluated CM using extensive physiologically realistic simulations and a dynamic [18F]FDG TB-PET dataset.
• Utilized a standard single-input two-tissue compartment model for analysis.

Main Results:

• CM achieves accuracy comparable to MCMC (median absolute percent error < 5%) but is over five orders of magnitude faster.
• CM provides reliable parametric images (e.g., K_i) without assuming irreversibility, unlike the Patlak method.
• The model generates valuable information on the statistical uncertainty of parameter estimates.

Conclusions:

• The generative consistency model (CM) removes computational barriers for Bayesian parametric imaging in TB-PET.
• CM offers a faster, more accurate, and statistically robust alternative to existing methods for whole-body kinetic analysis.
• The framework is adaptable to various tracers and compartment models, broadening its applicability.