Automated Multi-class Fetal Cardiac Vessel Segmentation in Aortic Arch Anomalies Using T2-Weighted 3D Fetal MRI

Abstract

Congenital heart disease (CHD) encompasses a range of cardiac malformations present from birth, representing the leading congenital diagnosis. 3D volumetric reconstructions of T2w black blood fetal MRI provide optimal vessel visualisation, supporting prenatal CHD diagnosis, a key step for optimal patient management. We present a framework for automated multi-class fetal vessel segmentation in the setting where binary manual labels of the vessels region of interest (ROI) are available for training, as well as a multi-class labelled atlas. We combine deep learning label propagation from multi-class labelled condition-specific atlases with 3D Attention U-Net segmentation to achieve the desired multi-class output. We train a single network to segment 12 fetal cardiac vessels for three distinct aortic arch anomalies (double aortic arch, right aortic arch, and suspected coarctation of the aorta). Our segmentation network is trained by combination of a multi-class loss, which uses the propagated multi-class labels; and a binary loss which uses binary labels generated by expert clinicians. Our proposed method outperforms label propagation in accuracy of vessel segmentation, while succeeding in segmenting the anomaly area of all three CHD diagnoses included, achieving a 100% vessel detection rate.