A 3D Dense-U-Net for fully automated 5D flow MRI segmentation

Abstract

Recently, a free-running 5D flow framework was introduced and validated. However, some 5D flow MRI is based on 3D radial imaging, which is limited by reduced SNR that can result in challenges with 3D segmentation. A number of previous studies have investigated automatic segmentation for 4D flow MRI, however these have been traditionally optimized for Cartesian datasets, which are typically acquired over much smaller spatial matrices and cover only one respiratory position. The purpose of this study was thus to adapt and expand a deep-learning framework to cardiac 5D flow MRI data for automatic segmentation of the thoracic aorta.