Bayesian Posterior Distribution Estimation of Kinetic Parameters in Dynamic Brain PET Using Generative Deep Learning...

Area of Science:

• Neuroimaging
• Medical Physics
• Computational Biology

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 dynamic PET data.
• Image noise introduces uncertainty in kinetic parameters, necessitating robust estimation methods like Bayesian inference.

Purpose of the Study:

• To develop a computationally efficient deep learning method for estimating kinetic parameter posterior distributions in dynamic PET imaging.
• To leverage an Improved Denoising Diffusion Probabilistic Model (iDDPM) for enhanced accuracy and speed compared to existing techniques.

Main Methods:

• Proposed an iDDPM-based approach for estimating kinetic parameter posterior distributions in dynamic PET.
• Evaluated the method on a [18F]MK6240 study, comparing it against Conditional Variational Autoencoder with dual decoder (CVAE-DD) and Wasserstein GAN with gradient penalty (WGAN-GP).
• Utilized Metropolis-Hasting Markov Chain Monte Carlo (MCMC) for reference posterior distribution inference.

Main Results:

• The iDDPM-based method demonstrated superior performance over CVAE-DD and WGAN-GP.
• Achieved significant computational speedup (over 230x faster) compared to MCMC methods.
• Inferred accurate kinetic parameter posterior distributions with low mean error (< 0.67%) and high precision (standard deviation error < 7.23%).

Conclusions:

• The proposed iDDPM method offers a computationally efficient and accurate alternative for kinetic parameter estimation in dynamic PET.
• This advancement holds promise for improving the analysis of neurodegenerative diseases like Alzheimer's using PET imaging.
• Deep learning models can effectively address noise challenges in PET imaging, enhancing diagnostic capabilities.