A fast 4D B-spline framework for model-based reconstruction and regularization in vector flow imaging

Abstract

A generic framework for model-based regularization and reconstruction is described, with applications in a wide range of noisy measurement scenarios. The framework employs automatic differentiation and stochastic gradient optimizers to perform online measurement fitting and regularization, and was implemented as a scalable CPU and GPU library with high-performance operation even in compute-or memory-intensive contexts, such as for 4D cardiac vector flow imaging. The framework was demonstrated by reconstructing 4D vector flow mapping through the incorporation of the incompressible Navier-Stokes equations. Furthermore, the achieved performance was within bedside applicability requirements.