Abstract

Cardiac MRI makes it possible to explore blood flow in three orthogonal directions within the cardiovascular system, through an acquisition sequence called 4D flow MRI. This sequence has been used in recent years to diagnose complex cardiovascular diseases with high accuracy. However, it is limited by the long times required to obtain an accurate three-dimensional segmentation of a region of interest that allows quantification of a series of hemodynamic parameters. Segmentation of these images is challenging due to problems such as low signal-to-noise ratio, phase accumulation errors in the images, spatiotemporal resolution, and respiratory motion. To address this challenge, we propose a processing pipeline that uses a neural network for medical image registration and a cascade of semantic segmentation neural networks. This approach improves the segmentation of pulmonary artery branches and aortic sections into single or multiple cardiac phases, facilitating clinical analysis and research in cardiovascular disease.