On the importance of appropriate fibre population selection in diffusion tractography

Introduction A great deal of recent research in diffusion MRI has focussed on improved modelling and reconstruction of multiple white matter fibre orientations within single image voxels. These methods are typically used to create an orientation distribution function (ODF) at each voxel, which represents the orientations of each fibre population crossing the voxel. However, very little attention has been given to the important problem of how a tractography algorithm should best navigate through the resulting field of ODFs. By far the commonest approach is to apply a continuity criterion, whereby a candidate propagation direction is selected whose orientation most closely matches the direction of entry to the voxel—but this assumption of smoothness dramatically oversimplifies the variety and complexity of tract shapes in the brain. “Global” approaches to tractography (e.g. [1,2]) offer a more sophisticated approach, but generally require substantially greater computation time. Here we propose a simple alternative strategy, based on prior knowledge, in which fibre population selection is based on the local orientation of a predefined reference tract. Methods Eight healthy young subjects (4 male, mean age 31.9 ± 5.3 yr) underwent a diffusion MRI protocol on a GE Signa LX 1.5 T clinical system. Echoplanar diffusion-weighted images were acquired along 64 noncollinear directions at a b-value of 1000 s mm, along with 7 b=0 images. Reconstructed image resolution was 2 x 2 x 2 mm. Scan time was approximately 20 min. ODFs were estimated using the Bayesian approach implemented in FSL, BEDPOSTX [3], with a maximum of two fibre populations at each voxel. Reference tracts were derived from a human white matter atlas as described in [4], represented as cubic B-splines, and transformed into the space of each subject along with the seed point. Probabilistic streamline tractography was performed by locally sampling from each fibre population—except where one population had less than a 5% estimated volume fraction—and choosing the sample whose orientation most closely matched the incoming direction (the continuity criterion) or the local orientation of the reference tract (the prior knowledge criterion). A small step of 0.5 mm was taken in the chosen direction, and then the process was repeated until the streamline was terminated due to leaving the brain or making a very sharp turn of approximately 80o or more. The probabilistic trilinear interpolation scheme described by Behrens et al. [3] was used to sample data at each step. 5000 streamlines were generated in this way, and a streamline visitation map calculated for each subject.