Reducing Geometric Uncertainty in Computational Hemodynamics by Deep Learning-Assisted Parallel-Chain MCMC.

Area of Science:

• Cardiovascular research
• Computational fluid dynamics
• Biomedical engineering

Background:

• Computational hemodynamic modeling is crucial for cardiovascular research and healthcare.
• Model reliability hinges on quantifying and reducing uncertainties in parameters and boundary conditions.
• Vascular geometry uncertainty significantly impacts hemodynamic simulation accuracy.

Purpose of the Study:

• To develop a Bayesian framework for propagating and reducing uncertainty in vascular geometries.
• To present a novel deep learning (DL)-assisted parallel Markov chain Monte Carlo (MCMC) method for efficient uncertainty quantification.
• To improve the accuracy and reliability of computational hemodynamic models.

Main Methods:

• A deep learning model was developed to approximate the geometry-to-hemodynamic relationship.
• An active learning strategy trained the DL model using online data from parallel MCMC chains.
• The DL model facilitated early rejection of improbable parameter proposals, accelerating convergence.

Main Results:

• The proposed DL-assisted MCMC method demonstrated efficient Bayesian posterior sampling.
• Geometric uncertainty was effectively reduced in two-dimensional aortic flow simulations.
• The method showed significant merit in improving the convergence and efficiency of uncertainty quantification.

Conclusions:

• The novel DL-assisted MCMC approach offers an efficient solution for uncertainty reduction in hemodynamic modeling.
• This method enhances the reliability of cardiovascular simulations by addressing geometric uncertainties.
• The findings support the integration of machine learning with Bayesian inference for complex biomedical modeling.