Automated White Matter Fiber Tract Segmentation for the Brainstem.

Abstract

This study aimed to develop an automatic segmentation method for brainstem fiber bundles. We utilized the brainstem as a seed region for probabilistic tractography based on multishell, multitissue constrained spherical deconvolution in 40 subjects from the Human Connectome Project (HCP). All tractography data were registered into a common space to construct a brainstem fiber cluster atlas. A total of 100 fiber clusters were identified and annotated. Cortical parcellation-based fiber selection was then performed to extract fibers within the annotated clusters that projected to their corresponding cortical regions. This atlas was applied for automatic brainstem fiber bundle segmentation in 10 HCP subjects and 8 patients with brainstem cavernous malformations. The spatial overlap between automatic and manual reconstruction was assessed. Ultimately, eight fiber bundles were identified in the brainstem atlas on the basis of their trajectories: the corticospinal tract (CST), corticobulbar tract, frontopontine tract, parieto-occipital-pontine tract, medial lemniscus, and superior, middle, and inferior cerebellar peduncles. The mean and standard deviation of the weighted dice (wDice) scores between the automatic and manual reconstructions were 0.9076 ± 0.0950 for the affected CST, 0.9388 ± 0.0439 for the contralateral CST, 0.9130 ± 0.0588 for the affected medial lemniscus, and 0.9600 ± 0.0243 for the contralateral medial lemniscus. This proposed method effectively distinguishes major brainstem fiber bundles across subjects while reducing labor costs and interoperator variability inherent to manual reconstruction. Additionally, this method is robust in that it allows for the visualization and identification of fiber tracts surrounding brainstem cavernous malformations.