A Variational Bayesian inference method for parametric imaging of PET data

Area of Science:

• Nuclear medicine
• Biomedical imaging
• Quantitative analysis

Background:

• Dynamic Positron Emission Tomography (PET) studies utilize compartmental models for detailed tracer kinetics.
• Voxel-level model inversion in PET is challenging due to low signal-to-noise ratio in time activity curves.

Purpose of the Study:

• To propose and validate a Variational Bayesian (VB) approach for robust quantitative parametric mapping in dynamic PET.
• To address non-uniform noise distributions and develop a hierarchical scheme for parameter priors.

Main Methods:

• Adaptation of Variational Bayesian (VB) methods to PET data with non-uniform noise.
• Implementation of a hierarchical scheme for deriving model parameter priors directly from images.
• Testing on synthetic data with compartmental models of increasing complexity.

Main Results:

• Accurate parameter estimates from synthetic data (bias < 2% ± 2%, RMSE < 15% ± 5%).
• Generation of reliable parametric maps from real PET data, even with high noise levels (unreliable estimates < 11% ± 5%).

Conclusions:

• The proposed VB approach effectively overcomes signal-to-noise limitations in dynamic PET.
• This method enables robust quantitative analysis and parametric mapping for novel PET tracers.