Input-Parameterized Physics-Informed Neural Networks for Time-Resolved 3D Blood Flow Velocity Reconstruction and Wall Shear Stress Calculation Using Data from Modified 2D PCMRI

Two-dimensional phase-contrast magnetic resonance imaging (2D PCMRI) quantifies velocity in a single imaging plane with unidirectional velocity encoding. To accurately measure velocity, the imaging plane must be positioned perpendicular to the blood vessel's centerline, requiring slice selection and adjusting the magnitude of bipolar gradients in the directional coils.

We propose a new method and processing algorithms to generate three-component volumetric velocities near the interrogation plane. Our method retains the slice selection scheme of traditional 2D PCMRI but introduces alternating, axis-aligned, velocity-sensitive bipolar gradient scans, reducing the temporal sampling frequency by one-third.

We use a deep learning framework called input-parameterized physics-informed neural networks (IP-PINN) to enhance the spatiotemporal resolution of velocity data. By leveraging fluid dynamics principles and MR physics, the IP-PINN algorithm accurately interpolates and extrapolates velocities, producing comprehensive 3D volumetric velocity fields and precise lumen boundary predictions around the imaging plane. These are crucial for calculating wall shear stress (WSS). This method is expected to improve 2D PCMRI's accuracy in calculating velocity-derived hemodynamic parameters.