Posterior estimation using deep learning: a simulation study of compartmental modeling in dynamic positron emission...

Area of Science:

• Medical Imaging
• Deep Learning
• Bayesian Inference

Background:

• Medical images are typically treated as deterministic, neglecting their inherent uncertainties.
• Underexplored uncertainties in medical imaging limit diagnostic accuracy and parameter estimation.

Purpose of the Study:

• To develop deep learning methods for efficient estimation of posterior distributions of imaging parameters.
• To derive accurate parameter estimates and their uncertainties from medical imaging data.

Main Methods:

• Implemented variational Bayesian inference using conditional variational auto-encoders (CVAE), including CVAE-dual-encoder and CVAE-dual-decoder architectures.
• Applied these deep learning models to a simulation study of dynamic brain Positron Emission Tomography (PET) imaging.

Main Results:

• The CVAE-dual-encoder and CVAE-dual-decoder models accurately estimated posterior distributions of PET kinetic parameters.
• Results showed good agreement with distributions sampled by Markov Chain Monte Carlo (MCMC).
• CVAE-vanilla demonstrated utility but with inferior performance compared to the dual-encoder/decoder models.

Conclusions:

• Deep learning approaches effectively estimate posterior distributions in dynamic brain PET imaging.
• The proposed methods align well with unbiased MCMC estimations, offering adaptable solutions for various applications.
• The choice of neural network architecture allows for user-specific application tuning.