Physics-informed neural networks for enhancing medical flow magnetic resonance imaging: Artifact correction and mean pressure and Reynolds stresses assimilation.

This article focuses on the application of physics-informed neural networks (PINNs) to enhance the accuracy of standard Cartesian four-dimensional (4D) flow magnetic resonance imaging (MRI) data in assessing turbulent mean flow fields through an in vitro axisymmetric stenosis. The study demonstrates that PINNs can effectively correct displacement artifacts and reduce noise in standard 4D flow MRI velocity measurements by enforcing the continuity equation, producing velocity fields comparable to those obtained from the artifact-free but time-consuming synchronized single-point imaging (Sync SPI) MRI. Subsequently, the PINN framework assimilates additional hemodynamic quantities—mean pressure and Reynolds stresses—using the Reynolds-averaged Navier–Stokes (RANS) equations closed with the Spalart–Allmaras turbulence model, validated against computational fluid dynamics (CFD) data and applied to experimental measurements. This two-step PINN approach offers a promising post-processing method to improve noninvasive cardiovascular flow assessments with short acquisition times, potentially aiding the diagnosis of conditions like aortic stenosis by providing continuous, physically consistent flow fields and derived biomarkers.