Pew-Thian Yap1, Yasheng Chen2,
Hongyu An2, John Gilmore3, Weili Lin2,
Dinggang Shen2
1Department of Radiology, University of
North Carolina at Chapel Hill, Chapel Hill, NC, United States; 2Department
of Radiology, University of North Carolina, Chapel Hill, NC, United States; 3Department
of Psychiatry, University of North Carolina, Chapel Hill, NC, United States
Fourth
Order Tensors (FOTs) give elegant mathematical properties akin to that of the
second order tensors. Recent formulation of FOT imposes positivity on the
estimates to ensure soundness in a physical sense - a property not often
found in more general higher order tensor approximations. Employing FOTs, we
propose a fast feature-based multi-scale registration algorithm for whole
brain HARDI data. Our registration algorithm requires a low computation cost
5 minutes to register a pair of 128x128x80 images at 2mm isotropic
resolution making it practically feasible for clinical applications. Our
methods involve three major components: 1) Generation of FOT-based features,
2) Hierarchical correspondence matching, 3) Dense deformation field
estimation, and 4) Retransformation.