Tradeoffs in Supersampling of DTI Metrics

Most tractography methods based on diffusion tensor images (DTIs) require repeated evaluation of tensors or tensor properties at locations not on the point lattice of measurements made during imaging [1]. Furthermore, computing tractwise statistics on scalar measures [2] may also require interpolation to such locations. The process of computing tensor measures comprises two stages: first computing the DTI from the diffusion-weighted images (DWIs), and second deriving the measures from the DTI. While it is appealing to interpolate the original DWIs at the desired point and then recompute the needed values, such an operation is computationally expensive, even for simple linear tensor-fitting algorithms. Interpolation at higher levels, however, is also problematic: Euclidean interpolation of tensor elements results in “tensor swelling” and may give singular results [3], while interpolating in the Riemannian manifold prevents these artifacts at great computational cost [4].