A powerful aspect of diffusion MR imaging is the ability to reconstruct fiber orientations in brain white matter; however, the application of traditional learning algorithms is challenging due to the directional nature of the data. In this paper, we present an algorithmic approach to clustering such spatial and orientation data and apply it to brain white matter supervoxel segmentation. This approach is an extension of the DP-means algorithm to support axial data, and we present its theoretical connection to probabilistic models, including the Gaussian and Watson distributions. We evaluate our method with the analysis of synthetic data and an application to diffusion tensor atlas segmentation. We find our approach to be efficient and effective for the automatic extraction of regions of interest that respect the structure of brain white matter. The resulting supervoxel segmentation could be used to map regional anatomical changes in clinical studies or serve as a domain for more complex modeling.