3D elastic registration improves HARDI-derived fiber alignment and automated tract clustering

High angular resolution diffusion imaging (HARDI) allows population studies of fiber integrity and connectivity. Tractography can extract individual fibers. For group studies, fibers must be clustered into recognizable bundles found consistently across subjects. Nonlinear image registration may improve population clustering. To test this, we performed whole-brain tractography with an orientation distribution function based Hough transform method in 20 young adults scanned with 4 Tesla, 105-gradient HARDI. We warped all extracted fibers to a geometrically-centered template using a 3D elastic registration driven by fractional anisotropy maps, to align embedded tracts. Fiber alignment was evaluated by calculating distances among corresponding fibers across subjects. Before and after warping, we performed spectral clustering of the fibers using a k-means method, based on eigenvectors of a fiber similarity matrix. In tests with an overlap metric, non-rigid fiber warping yielded more robust clustering results. Non-rigid warping is therefore advantageous for population studies using multi-subject tract clustering.