Magnetic Resonance Imaging-Based 4D Flow: The Role of Artificial Intelligence

With 4D flow MRI (time-resolved three-dimensional phase contrast MRI) the velocity of blood flow can be measured in three directions, over several time frames of the cardiac cycle and in a volume comprising the heart, aorta and pulmonary vasculature. The technique has shown clinical relevance in a wide variety of cardiothoracic disease e.g., cardiomyopathy, congenital heart disease, aortic and pulmonary disease. The drawbacks of 4D flow MRI are long scan times and complex processing of the data. The acquisition can be accelerated by acquiring less data (undersampling) and reconstructing the images with advanced mathematics. But reconstruction times are often long, which hampers clinical implementation. To extract biomarkers from the data that are relevant for disease, often complex and time-consuming manual segmentations are needed, further limiting the clinical application. To improve on reconstruction and segmentation times, machine learning can provide fast and robust alternatives by networks trained on fully- and undersampled images and by networks classifying and locating tissue boundaries based on manual annotations. Furthermore, the image quality of 4D flow MRI (noise and spatio-temporal resolution) can be improved by networks trained on fluid dynamics simulations or by networks using regularizations based on fluid flow physics (physics informed networks).
In this chapter we review the basics of 4D flow MRI, we review the most recent developments in machine learning applications for 4D flow MRI and describe the clinical relevance of 4D flow MRI in a wide variety of cardiothoracic disease.