A Deep Learning Approach for Segmenting the Subplate and Proliferative Zones in Fetal Brain MRI

Abstract

In vivo fetal brain MRI is employed in clinical practice and in research studies to appreciate in utero brain development. There is increasing interest in transient regions of the fetal brain, such as the subplate (SP), ventricular zone (VZ) and ganglionic eminence (GE) (also referred to as germinal matrix), and their role in normal and abnormal antenatal brain development. On T1w and T2w fetal MRI, these transient regions are defined by highly heterogeneous and stratified signal intensities with rapidly changing patterns. In this work, we define the SP, VZ and GE in a 0.5mm isotropic resolution atlas from the developing Human Connectome Project (dHCP) [1, 17, 18], and train an Attention U-Net [12] to automatically segment them based on semi-automatically generated labels. Our solution spans from 21 through to 36 weeks gestational age (GA), offering insight into a crucial period of antenatal brain development. The proposed automated segmentation achieved mean Dice scores of 0.88, 0.70 and 0.82 for SP, VZ and GE respectively. A volumetric comparison of transient regions in a small cohort of fetuses with isolated ventriculomegaly (VM, n = 8) vs. controls (n = 265) showed significantly enlarged absolute volumes in the GE (P = 0.005) and VZ (P < 0.001) of the left hemisphere.