Supervised classification of white matter fibers based on neighborhood fiber orientation distributions using an ensemble of neural networks

White matter fibers constitute the main information transfer network of the brain and their accurate digital representation and classification is an important goal of neuroscience image computing. In current clinical practice, the reconstruction of desired fibers generally involves manual selection of regions of interest by an expert, which is time-consuming and subject to user bias, expertise and fatigue. Hence, automation of the process is desired. To that end, we propose a supervised classification approach that utilizes an ensemble of neural networks. Each streamline is represented by the fiber orientation distributions in its neighborhood, while the resolved fiber orientations are obtained by generalized q-sampling imaging (GQI) and a subsequent diffusion decomposition method. In order to make the supervised fiber classification succeed in a real scenario where a substantial portion of reconstructed fiber tracts contain spurious fibers, we present a way to create an “invalid” class label through a dedicated training set creation scheme with an ensemble of networks. The performance of the proposed classification method is demonstrated on major fiber pathways in the brainstem. 30 subjects from Human Connectome Project (HCP)’s publicly available “WU-Minn 500 Subjects + MEG2 dataset” are used as the dataset