Modeling Fetal Brain Development: A semi-automated platform for localization, reconstruction, and segmentation of the fetal brain on MRI
Abstract
With advances in fetal magnetic resonance imaging (MRI), research in neonatal neuroscience has shifted to identify in utero brain-based biomarkers for outcome prediction in high-risk fetuses, particularly those impacted by growth restriction. Volumetric segmentation of the fetal brain can provide better understanding of the trajectories of brain development and may aid in predicting functional outcomes. The current thesis aimed to develop semi-automatic methods to target deep brain structures in the fetal brain identified on MR images in fetuses with and without growth restriction. In this study, pregnant women (35-39 weeks gestational age [n=9]) with growth appropriate (n=8) and growth restricted fetuses (n=1) were recruited. Fetal MRI was performed on 1.5 Tesla (T) and 3T MRI scanners and 2D stacks of T2-weighted images were acquired. A novel fetal whole brain segmentation algorithm developed for second trimester fetuses was applied to the T2-weighted MR images to reconstruct 2D volumes into 3D images. To segment deep brain structures, an atlas of cortical and subcortical structures was registered to the 3D reconstructed images. Linear and nonlinear registration algorithms, with two types of similarity metrics (mutual information [MI], cross-correlation [CC]), were compared to determine the optimal strategy of segmenting subcortical structures. Dice coefficients were calculated to validate the reliability of automatic methods and to compare the performance between the registration algorithms compared to manual segmentations. Comparing atlas-generated masks against manually segmented masks of the same brain structures, the median Dice-kappa coefficients for linear registration using CC performed optimally. However, post hoc analyses indicated that linear MI and CC performed comparably. Overall, this semi-automatic subcortical segmentation method for third-trimester fetal brain images provides reliable performance. This segmentation pipeline can aid in identifying early predictors of brain dysmaturation to support clinical decision making for antenatal treatment strategies and promote optimal neurodevelopment in fetuses.