Spectral clustering based parcellation of FETAL brain MRI

Abstract

Many neuroimaging studies are based on the idea that there are distinct brain regions that are functionally or micro-anatomically homogeneous. Obtaining such regions in an automatic way is a challenging task for fetal data due to the lack of strong and consistent anatomical features at the early stages of brain development. In this paper we propose the use of an automatic approach for parcellating fetal cerebral hemispheric surfaces into K regions via spectral clustering. Unlike previous methods, our technique has the crucial advantage of only relying on intrinsic geometrical properties of the cortical surface and thus being unsupervised. Results on a data-set of fetal brain MRI acquired in utero demonstrated a convincing parcellation reproducibility of the cortical surfaces across fetuses with varying gestational ages and folding magnitude.