Advancing Time-Resolved Phase-Contrast MRI Analysis: Development and Application of the Input-Parameterized Physics-Informed Neural Network (IP-PINN)

Abstract

Time-Resolved Three-Dimensional Phase-Contrast MRI (4D Flow MRI) is a powerful non-invasive technique for quantitatively assessing cardiovascular hemodynamics. Despite its potential, the clinical application of 4D Flow MRI is constrained by coarse spatio-temporal resolution, acquisition noise, and artifacts including velocity aliasing and eddy current induced phase offsets. These limitations compromise the accuracy and reliability of hemodynamic assessments, particularly in complex vascular structures. This thesis proposes a novel deep learning-based framework called the Input-Parameterized Physics Informed Neural Network (IP-PINN) to address these challenges. By integrating advanced machine learning with the underlying physics of blood flow, the IP-PINN framework enhances low-resolution 4D Flow MRI data, attenuates acquisition noise, and mitigates velocity aliasing and phase offset artifacts. The framework leverages a ResNet-based convolutional neural network to encode input data into a latent vector, which is then utilized by a feedforward neural network to produce a continuous spatio-temporal representation of the variables of interest. The IP-PINN’s unique ability to generalize across different datasets by parameterizing the solution with respect to the input velocity-encoded images significantly reduces the need for time-intensive retraining. The IP-PINN does not require ground-truth lables and pre-training with either real low resolution image data or synthetic data from computational fluid dynamics (CFD) simulations enhances the framework's applicability. The IP-PINN preserves the continuous spatio-temporal representation and the ability to generate truncation error-free derivatives characteristic of PINNs, while significantly expediting processing time. Benchmark tests against simulated datasets demonstrate that IP-PINN is over an order of magnitude faster than traditional PINNs, achieving superior accuracy. Additionally, the method generates high-resolution magnitude images for lumen boundary segmentation, relying solely on velocity-encoded scans and negating the need for reference scans. Building on an initial implementation that operated in complex image space and relied on the three velocity-encoded scans, the thesis extends IP-PINNs in two directions. First, reconstruction of three-component, three-dimensional (3D-3C) velocity fields and high-resolution spin-density maps in the vicinity of the imaging plane using data from minimally altered 2D PC-MRI sequence. Second, reconstruction of 3D-3C velocity maps and spin-denisty maps from a pseudo one-point 4D-Flow MRI sequence, that acquires only one velocity-encoded dataset per slice, reducing the raw data burden by 75%. For this extremely sparse regime, the data-fidelity term is reformulated directly in k-space, which preserves the exact acquisition physics, avoids the convolution blurring inherent in image-space. With a dramatically reduced execution time of approximately two minutes, operating on undersampled acquired data, and simplified operational requirements (no need for specifying geometry and boundary conditions), the IP-PINN promises to advance the state-of-the-art in hemodynamic assessment, offering a robust and efficient solution for enhancing time-resoleved PCMRI data, with significant implications for both clinical practice and cardiovascular research.