Bayesian finite element regression for vascular flow reconstruction with quantified uncertainty

Abstract

Reconstructing accurate velocity and pressure fields from under-resolved noisy measurements of blood flow is an ill-posed inverse problem due to unknown inlet and outlet boundary conditions. We present a Bayesian finite element regression framework that reconstructs steady three-dimensional velocity and pressure fields, with quantified uncertainty, from noisy velocity observations without offline training data. We represent velocity and pressure fields in Taylor-Hood finite element basis functions, and construct physics-informed priors on the nodal degrees of freedom from maximum-entropy principles. Combined with a likelihood specified by a noise-model, this yields a posterior whose maximum-a-posteriori estimate (MAP) gives velocity and pressure reconstructions. The MAP estimate is computed by solving a large-scale sparse nonlinear least-squares problem where pressure is eliminated analytically, no-slip walls are enforced exactly, and gradient is computed without forward/adjoint solves or automatic differentiation. A Laplace approximation of the posterior quantifies the uncertainties in our reconstructions and propagates them to clinically relevant quantities of interest including, pressure drop, flow rates, and wall shear stress. On patient-specific cerebral aneurysm, aortic aneurysm, and aortic coarctation geometries, the method reconstructs velocity and pressure more accurately than tricubic interpolation and comparably to a PINN, while recovering region-of-interest wall shear stress more accurately than both.