Fiber-distance-based unsupervised clustering of MR tractography data

Abstract

Background: MR tractography from diffusion tensor imaging provides a non-invasive way to explore white matter pathways in the human brain. However, a challenge to extracting reliable anatomical information from these data is the use of reliable and effective clustering methodologies. In this paper, we implemented a new version of a robust unsupervised clustering method from MR tractography data using the density-based spatial clustering of applications with noise (DBSCAN) algorithm. New method: Conventional DBSCAN clustering methods for MR tractography data use each fiber's start and end point as well as the distance between start and end points. Instead, in this study, we extracted and used a fiber-distance matrix generated for all fiber combinations from the tractography dataset in DBSCAN clustering. The two DBSCAN parameters-minimum point number and maximum radius of the neighborhood-were selected according to the value generated with the cluster stability index (CSI). Results: Performing the proposed CSI-optimized DBSCAN-based clustering method on MR tractography data of the superior longitudinal fasciculus generated 6 robust, non-overlapping, clusters that are neuroanatomically related. Comparison with existing methods: Conventional DBSCAN-based clustering methods have intrinsic error potential in the clustering results due to deviations in fiber shape and fiber location. The proposed method did not exhibit clustering error caused by deviation in fiber trajectory or fiber location. Conclusions: We implemented a new, robust DBSCAN-based fiber clustering method for MR tractography data. The CSI-optimized DBSCAN-based unsupervised clustering is applicable to investigation of the neuroconnectome and the fiber structure of the brain.