Free-Spin Tracking: A Novel Global Tractography Algorithm

Diffusion MRI [1] and fiber tractography have become a reference to study the neural structure of the brain white matter. Standard tractography algorithms behave generally locally, i.e., a single fiber is initiated at a seed point and is constructed iteratively, thus making fibers subject to cumulative errors. On the contrary, in global approaches the entire white matter is tracked at once: in case of an ambiguity within a voxel, several fiber contributions coming from the entire volume can help choose the correct neural pathway. Global approaches have been proposed recently, such as the Gibbs tracking [2], which is an analogy to polymer creation in chemistry: pieces of fibers arrange themselves to minimize the resemblance between the simulated and real diffusion-weighted images (DWI). Although such approach is a very promising alternative to the standard techniques, it appears to be computationally expensive and uses the raw DWI instead of a diffusion model like the tensor [1]. Raw DWI are subject to MR noise, and model fitting can help remove this noise [3] and yields a good compression of the signal, which is important to speed up computation. Another global approach was proposed by Cointepas et al. [4], where fibers are obtained by optimizing a spin field. A spin represents a piece of fiber of unit length which tends to align along the more probable fiber direction (e.g., the main tensor eigenvector), while associating with other spins such that the curvature of the fiber joining them is minimal. Such approach is appealing because: 1. it relies on the only prior that brain fibers have a low curvature, and 2. each fiber can be optimized independently from the others making the parallelization of the algorithm straightforward. However, the proposed model has severe limitations: only one spin per voxel was placed, which assumes only one fiber contribution per voxel (which is not true inside crossings), and spin position was fixed which may lead to a non-optimal sampling of the fibers. In this work, we propose to liberate the spins so that they are free to move, and that their number within a region (like a voxel) is driven by the data by allowing the creation, or birth, of new spins. We show that even diffusion tensors can be used to successfully reconstruct fiber crossings and kissings at a reasonable computational cost.