Recent work (2004-2007) at the Harvard LMI lab has shown that spectral clustering can be successfully used to create bundles from DTI fiber tracts. These clusters do not always correspond to anatomical bundles but they do to a large extent.
The picture on the right shows DTI fiber tracts extracted from the corpus callosum.
Some of my results of spectral clustering, applied to this data set, are presented below.
The clusters are consistent across 12 same-sex subjects--we get bundles from the rostrum, genu, corpus and splenium, the 4 sections of the corpus callosum, in each case. The distance measure is the mean closest point described in this paper by O'Donnell and Westin. Interestingly, a median closest point distance failed to produce more than one cluster--i.e. the distances between individual fibers were all the same in this case.
