Machine Learning-Based Denoising and Imputation of Unsteady Cardiovascular Flow Data

Obtaining clean, high resolution velocity measurements of blood flow inside small arteries such as cerebral vasculature is challenging, as the existing experimental techniques have several limitations. While time-resolved three-dimensional phase-contrast magnetic resonance imaging (4D-Flow MRI) and particle image velocimetry (PIV) are the most common approaches in clinical and laboratory settings, respectively, they are constrained by low spatio-temporal resolution, noise, and other artifacts. Therefore, handling corrupt blood flow data is a key challenge towards developing more accurate and robust cardiovascular flow models. There are well developed algorithms in the machine learning community that can tackle similar issues, such as data imputation, denoising, or outlier detection. Existing methods have been less frequently used and leveraged for complex real-world fluid flow problems, such as cardiovascular flows. This study investigates and compares several techniques for filling in missing values and denoising unsteady hemodynamics data. Most algorithms for such tasks are based on singular value decomposition (SVD) or autoencoders (deep learning). These methods are tested on voxel-based data mimicking certain 4D-Flow MRI data features created from computational fluid dynamics (CFD) simulations, and patient-specific in-vitro 4D-Flow MRI data. Different methods are compared, and the associated challenges are addressed for synthetic and real-world experimental data.