We propose a fiber clustering algorithm composed by several steps, with the objective of representing the whole dataset by a small set of cluster centroids. First, a clustering is performed separately for a subset of points within the streamlines. The obtained point clusters are then used to regroup the fibers having common point clusters. Next, fiber clusters are filtered out by size and finally regroup using a quick merge based on a maximum Euclidean distance. A reduced set of regular and thin clusters is finally obtained. In contrast to previous works, the proposed method is only based on streamline structure.