Multi-Label Segmentation White Matter Structures: Application to Neonatal Brains.

Abstract

Accurate and consistent segmentation of brain white matter bundles at neonatal stage plays an important role in understanding brain development and detecting white matter abnormalities for the prediction of psychiatric disorders. Due to the complexity of white matter anatomy and the spatial resolution of diffusion-weighted MR imaging, multiple fiber bundles can pass through one voxel. The goal of this study is to assign one or multiple anatomical labels of white matter bundles to each voxel to reflect complex white matter anatomy of the neonatal brain. For this, we develop a supervised multi-label k-nearest neighbor (ML-kNN) classification algorithm in Riemannian diffusion tensor spaces. Our ML-kNN considers diffusion tensors lying on the Log-Euclidean Riemannian manifold of symmetric positive definite (SPD) matrices and their corresponding vector space as feature space. The ML-kNN utilizes the maximum a posteriori (MAP) principle to make the prediction of white matter labels by reasoning with the labeling information derived from the neighbors without assuming any probabilistic distribution of the features. We show that our approach automatically learns the number of white matter bundles at a location and provides anatomical annotation of the neonatal white matter. In addition, our approach also provides the binary mask for individual white matter bundles to facilitate tract-based statistical analysis in clinical studies. We apply this method to automatically segment 13 white matter bundles of the neonatal brain and examine the segmentation accuracy against semi-manual labels derived from tractography.