A comparative study of 16 tractography algorithms for the corticospinal tract: reproducibility and subject-specificity.

Methods– Data acquisition. We scanned 9 healthy subjects (3F/6M, 31.25 ± 4.2y) at three time points separated by a 2 weeks interval. Diffusion-weighted EPI were acquired on a Siemens 3T Verio scanner, b = 1000 s/mm2, on 64 evenly distributed directions plus 2 b = 0 images. Pre-processing. The diffusion-weighted images were regularized and corrected for Rician noise. The diffusion tensor was estimated using linear least squares fitting, and fractional anisotropy (FA) maps were computed. Deformation field from and to the JHU Eve template were computed using non-linear registration of the FA images. Fiber tracking. The fiber tracking of left and right corticospinal tracks (CST) was performed in subject-space. When applicable, the tractography was initiated in the cerebral peduncle (CP), using 5 seeds per voxel in that region of interest (ROI). The step size was set to 0.2mm, and the maximum turning angle to 35°. Tracking was terminated whenever underlying FA fell below 0.15. The tract was a posteriori filtered to keep only streamlines going through the posterior limb of the internal capsule (PLIC), and decimated to keep only 300 streamlines per track. The regions of interest were taken from Eve labels, warped to subject space using the above-computed deformation field. Tracking algorithms are streamline and probabilistic streamline on constrained spherical deconvolution (MRtrix), FACT, second-order Runge-Kutta, streamline and tensorline on tensor (Diffusion toolkit), Gibbs globals tracking (MITK), 4th order Runge-Kutta and tensorline on tensor (Camino), tensorline, streamline, tensor deflection, bootstrap streamline, bootstrap tensor deflection on tensor and streamline on CSA ODF (in-house software), and unscented Kalman filter on multi-tensor model. Statistical analysis. Tract similarity was computed using a Gaussian process framework. To this end, normalized tract probability images were computed in subject space, and warped back to template space. As a summary of the similarity matrix, we computed the ratio of the average inter-subject distance over the intra-subject distances.