White matter fiber bundles of the human brain form a spatial pattern defined by the anatomical and functional architecture. Human brain atlases provide names for individual tracts and document that these patterns are comparable across subjects. Tractography applied to the tensor field in diffusion tensor imaging (DTI) results in sets of streamlines which can be associated with major fiber tracts. Comparison of fiber tract properties across subjects requires comparison at corresponding anatomical locations. As an alternative to linear and nonlinear registration of DTI images and voxel-based analysis, we propose a novel methodology that models the shape of white matter tracts. A clustering uses similarity of adjacent curves and an iterative processing scheme to group sets of curves to bundles and to reject outliers. Unlike previous work which models fiber tracts as sets of curves centered around a spine, we extend the notion of bundling towards a more general representation of manifolds. We describe tracts, represented as sets of curves of similar shape, by a shape prototype swept along a space trajectory. This approach can naturally describe white matter structures observed either as bundles dispersing towards the cortex or tracts defined as dense patterns of parallel fibers forming manifolds. Curves are parameterized by arc-length and represented by intrinsic local shape properties (curvature and torsion). Feasibility is demonstrated by modeling the left and right cortico-spinal tracts and a part of the transversal callosal tract.