Abstract:The segmentation of the fetal cerebral cortex from magnetic resonance imaging (MRI) is an important tool for neurobiological research about the developing human brain. Manual segmentation is difficult and time-consuming. Limited image resolution and partial volume effects introduce errors and labeling noise when attempting to automate the process through machine learning. The significant morphological changes observed during brain growth pose additional challenges for learning-based image segmentation methods, which may drastically increase the amount of necessary training data. In this paper, we propose a framework to learn from noisy labels by using additional regularization via shape priors for the accurate segmentation of the cortical gray matter (CGM) in 3D. Firstly, we introduce a novel structure consistency loss based on persistent homology analysis of the cortical topology. Secondly, a regularization loss term is proposed by integrating assumptions about the cortical thickness within each sample. Our experiments on the developing human connectome project (dHCP) dataset show that our method can predict accurate CGM segmentation learned from noisy labels.

Fetal Cortex Segmentation with Topology and Thickness Loss Constraints