Enhanced Automatic Segmentation for Superficial White Matter Fiber Bundles for Probabilistic Tractography Datasets

This paper presents an enhanced algorithm for automatic segmentation of superficial white matter (SWM) bundles from probabilistic dMRI tractography datasets, based on a multi-subject bundle atlas. Previous segmentation methods use the maximum Euclidean distance between corresponding points of the subject fibers and the atlas centroids. However, this scheme might include noisy fibers. Here, we propose a three step approach to discard noisy fibers improving the identification of fibers. The first step applies a fiber clustering and the segmentation is performed between the centroids of the clusters and the atlas centroids. This step removes outliers and enables a better identification of fibers with similar shapes. The second step applies a fiber filter based on two different fiber similarities. One is the Symmetrized Segment-Path Distance (SSPD) over 2D ISOMAP and the other is an adapted version of SSPD for 3D space. The last step eliminates noisy fibers by removing those that connect regions that are far from the main atlas bundle connections. We perform an experimental evaluation using ten subjects of the Human Connectome (HCP) database. The evaluation only considers the bundles connecting precentral and postcentral gyri, with a total of seven bundles per hemisphere. For comparison, the bundles of the ten subjects were manually segmented. Bundles segmented with our method were evaluated in terms of similarity to manually segmented bundles and the final number of fibers. The results show that our approach obtains bundles with a higher similarity score than the state-of-the-art method and maintains a similar number of fibers.Clinical relevance—Many brain pathologies or disorders can occur in specific regions of the SWM automatic segmentation of reliable SWM bundles would help applications to clinical research.