A Novel Similarity Measure for Fiber Clustering using Longest Common Subsequence Christian Böhm University of Munich boehm@dbs.ifi.lmu.de Jing Feng University of Munich feng@dbs.ifi.lmu.de Xiao He University of Munich he@dbs.ifi.lmu.de Son T. Mai University of Munich mtson@dbs.ifi.lmu.de Claudia Plant Florida State University cplant@fsu.edu Junming Shao University of Munich shao@dbs.ifi.lmu.de ABSTRACT Diffusion tensor imaging (DTI) is an MRI-based technology in neuroscience which provides a non-invasive way to explore the white matter fiber tracks in the human brain. From DTI, thousands of fibers can be extracted, and thus need to be clustered automatically into anatomically meaningful bun- dles for further use. In this paper, we focus on the essential question how to provide an efficient and effective similarity measure for the fiber clustering problem. Our novel sim- ilarity measure is based on the adapted Longest Common Subsequence method to measure shape similarity between fibers. Moreover, the distance between start and end points of a pair of fibers is also included with the shape similarity to form a unified and flexible fiber similarity measure which can effectively capture the similarity between fibers in the same bundles even in noisy conditions. To enhance the efficiency, the lower bounding technique is used to restrict the compar- ison of two fibers thus saving computational cost. Our new similarity measure is used together with density-based clus- tering algorithm to segment fibers into groups. Experiments on synthetic and real data sets show the efficiency and effec- tiveness of our approach compared to other distance-based techniques, namely Dynamic Time Warping (DTW), Mean of Closest Point (MCP) and Hausdorf (HDD) distance. Categories and Subject Descriptors H.2.8 [Database Applications]: Data mining General Terms Algorithms Keywords Fiber Clustering, Longest Common Subsequence, Diffusion Tensor Imaging 1. INTRODUCTION Diffusion tensor imaging (DTI) is a structural magnetic resonance imaging (MRI) which helps to measure micro- scopic movement of water in the brain, and it has been used to explore organization and integrity of white matter struc- tures of human brain in vivo. From DTI data, the white matter tracts can be reconstructed via a process called trac- tography which estimates white matter tract trajectories by following likely track directions [13]. This information can be used in surgical planning and in study anatomical con- nectivity, brain changes and mental disorders [12]. After tractography, we obtain thousands of fiber trajec- tories, and they need to be grouped into anatomical mean- ingful structures before being used. One common method is based on the knowledge of experts and is referred to as vir- tual dissection [4]. This method interactively selects fibers passing through some manually defined region of interests (ROIs). This process is highly flexible and can help to detect anatomically meaningful bundles with very different shapes. However, it is very time consuming and may be biased by subjective points of view of the experts. Therefore, auto- matic clustering of white matter fiber tracks is an interesting alternative for many applications [5, 6, 3]. One essential problem in automatic fiber clustering is to provide a similarity measure for a pair of fibers. Two fibers are usually considered as similar if they are separated by a small distance, have comparable length and similar shape [6]. However, these criteria may be not sufficient. Two fibers with different shapes, for example, can be grouped into the same bundle if they start and end at the same regions [3]. Moreover, due to the scanning process of DTI, each fiber may contain some amount of noise, which can affect the similarity between them [4]. Although, there are many pro- posed techniques in literature [5, 14, 6, 15, 11], much efforts are going on to find out more effective and efficient proce- dures. Among various techniques, distance based similarity mea- sure ones like Hausdorf distance (HDD), Mean of Closest Point distance (MCP) [5] and Dynamic Time Warping (DTW) [14] are widely used. However, their point-to-point distance measure mechanism is sensitive to noise, which affects the final distance similarity between pairs of fibers. Moreover, only the final distance between two fibers is not enough to tell whether they have a similar shape or they are separated by a small distance. Thus, this limits their ability to effec- tively group fibers into meaningful bundles. Besides, with Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. DMMH’11, August 21, 2011, San Diego, California, USA. Copyright 2011 ACM 978-1-4503-0843-4//11/08...$10.00. 1