Tractography Gone Wild : Probabilistic Tracking Using the Wild Bootstrap

D. K. Jones Centre for Neuroimaging Sciences, Institute of Psychiatry, London, United Kingdom INTRODUCTION: DT-MRI is inherently a noisy technique, leading to uncertainty in estimates of the eigensystem of the diffusion tensor. Many tractography algorithms (particularly streamline approaches, e.g. 3,4) do not account for this uncertainty and thus provide no indication of the confidence one can assign to reconstructed pathways. Probabilistic approaches attempt to assign such a confidence, but typically do so by making a priori assumptions about the data. In contrast, bootstrap DT-MRI derives the uncertainty from the data itself without making any assumptions and has been used previously to derive uncertainty in the tensor eigensystem. Bootstrapping has been combined with tractography to create confidence maps for fiber trajectories and to examine distributions of parameters (principal eigenvector, eigenvalues, FA) at each vertex of a streamline as a way of identifying a range of artefacts. While the limited assumptions of the bootstrap make it an attractive option, the time required to collect sufficient data for accurate and precise bootstrapping can be prohibitive ̧leading to reduced subject compliance and increased movement artefact. Whitcher et al. recently proposed an alternative approach to deriving distributions of tensor-derived parameters using the Wild Bootstrap. In brief, this approach obtains probability distributions for model-parameters by permuting the residuals to the fitted model and refitting the model (see ref. 11 for full details). The huge advantage over previously implemented bootstrap methods is that it does not require the collection of extra data, and therefore brings the technique into the clinical realm. Whitcher et al. have shown that results obtained with the wild bootstrap are comparable with the ‘conventional’ bootstrap when considering data on a voxel-by-voxel basis. Here we combine the Wild Bootstrap with tractography and show ‘probabilistic’ tracking results that are comparable with those from the (much more time-consuming) conventional bootstrap.