A new combined distance measure for the clustering of fiber tracts in Diffusion Tensor Imaging ( DTI )

Introduction In recent years various fiber tractography methods have been evolved. Although these resulting tractograms offers plenty of information, they are rarely used in clinical routine, due to the fact that processing is often time-consuming and an experienced operator is essential to obtain good results. Apart from that, tractograms can be very useful for surgeons who need to know where the main fiber bundles are located and if they infiltrated or relocated by a tumor. Furthermore, tractograms can also be employed by researchers for studying neurological diseases such as Schizophrenia or Obsessive Compulsive Disorder (OCD). For example, the connections or even the degree of connectivity between certain regions in the brain may contain important information about the disease itself and is maybe beneficial for an early diagnosis or a proper treatment. The analysis of tractograms is usually performed manually by experienced users, who select a Region Of Interest (ROI) to define the tracts of interest (those tracts that cross the ROI). This however, depends strongly on the user and his level of experience and is by no means an objective measure. Therefore it is not reproducible and prone to errors. To overcome this limitations cluster analysis can be employed to partition fiber tracts into clusters with the final intention to minimize the intra-cluster differences (the difference between the fiber tracts in the cluster) and to maximize the inter-cluster variability (the difference between the clusters). Thereby, the similarity between individual tracts is determined through comparison of tract-specific features or similarity measures. In recent studies several similarity measures such as the Hausdorff distance, the corresponding segment ratio [1] or the Euclidean distance have been observed, but no optimal measure was found yet. All methods have deficiencies and the combination of appropriate and distinguishable similarity measures is recommended to improve the results [2]. Therefore, the aim of this study was to develop a new combined similarity measure that incorporates various features.