Combining Time Series Similarity with Density-Based Clustering to Identify Fiber Bundles in the Human Brain

• Published in: 2010 IEEE International Conference on Data Mining Workshops pp. 747 - 754
• Main Authors: Junming Shao, Hahn, Klaus, Qinli Yang, Böhm, Christian, Wohlschläger, Afra, Myers, Nicholas, Plant, Claudia
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2010
• Subjects: Clustering algorithms; Density-based Fiber Clustering; Diffusion Tensor Imaging; Dynamic Time Warping; Extraterrestrial measurements; Humans; Lower Bounding Distance; Noise; Shape; Time measurement; Time series analysis
• ISBN: 9781424492442; 1424492440
• ISSN: 2375-9232
• DOI: 10.1109/ICDMW.2010.15

Understanding the connectome of the human brain is a major challenge in neuroscience. Discovering the wiring and the major cables of the brain is essential for a better understanding of brain function. Diffusion Tensor imaging (DTI) provides the potential way of exploring the organization of white matter fiber tracts in human subjects in a non-invasive way. However, it is a long way from the approximately one million voxels of a raw DT image to utilizable knowledge. After preprocessing including registration and motion correction, fiber tracking approaches extract thousands of fibers from diffusion weighted images. In this paper, we focus on the question how we can identify meaningful groups of fiber tracks which represent the major cables of the brain. We combine ideas from time series mining with density-based clustering to a novel framework for effective and efficient fiber clustering. We first introduce a novel fiber similarity measure based on dynamic time warping. This fiber warping measure successfully captures local similarity among fibers belonging to a common bundle but having different start and end points. A lower bound on this fiber warping measure speeds up computation. The result of fiber tracking often contains imperfect fibers and outliers. Therefore, we combine fiber warping with an outlier-robust density-based clustering algorithm. Extensive experiments on synthetic data and real data demonstrate the effectiveness and efficiency of our approach.