Data-driven physics-informed blood-vessel boundary reconstruction from 4D Flow MRI

Accurate computation of hemodynamic indicators—such as wall shear stress (WSS)—is critical for cardiovascular disease prognosis. While 4D Flow MRI provides in-vivo blood velocity fields, its limited spatio-temporal resolution hinders precise hemodynamic assessment. Additionally, vessel wall segmentation from magnitude images is challenging, making WSS estimation difficult using 4D Flow MRI alone.

We previously introduced input-parameterized physics-informed neural networks (IP-PINNs) to super-resolve 4D Flow MRI and correct imaging artifacts. This work extends that framework to accurately estimate blood vessel boundaries. We represent velocity, pressure, magnitude (interpreted as apparent spin density), and porosity using multi-layer perceptrons (MLPs). Data fidelity is enforced by matching predicted and acquired k-space data for each velocity scan. Regularization includes total variation minimization of porosity, a binarization penalty, and constraints from the divergence-free condition and Brinkman-penalized Navier–Stokes equations. The method is validated on synthetic 4D Flow MRI data from CFD simulations of pulsatile flow in realistic aneurysm geometries.

This approach improves vessel boundary detection and enhances hemodynamic analysis from low-resolution 4D Flow MRI.