Identifying white-matter fiber bundles in DTI data using an automated proximity-based fiber-clustering method

We present a method for clustering diffusion tensor imaging (DTI) integral curves into anatomically plausible bundles. An expert rater evaluated the anatomical accuracy of the bundles. We also evaluated the method by applying an experimental cross-subject labeling method to the clustering results. We first employ a sampling and culling strategy for generating DTI integral curves and then constrain the curves so that they terminate in gray matter. We then employ a clustering method based on a proximity measure calculated between every pair of curves. We interactively selected a proximity threshold to achieve visually optimal clustering in models from four DTI datasets. An expert rater then assigned a confidence rating about bundle presence and accuracy for each of 12 target fiber bundles of varying calibers and type in each dataset. We then created a fiber bundle template to cluster and label the fiber bundles automatically in new datasets. According to expert evaluation, the automated proximity-based clustering and labeling algorithm consistently yields anatomically plausible fiber bundles on large and coherent clusters. This work has the potential to provide an automatic and robust way to find and study neural fiber bundles within DTI.