Pew-Thian Yap1, Yasheng Chen1,
Hongyu An1, John H. Gilmore2, Weili Lin1,
Dinggang Shen1
1Department of Radiology, University of
North Carolina, Chapel Hill, NC, United States; 2Department of
Psychiatry, University of North Carolina, Chapel Hill, NC, United States
We
propose a full-brain multi-scale feature-based deformable registration
algorithm based on the statistics of the diffusion profile of HARDI data.
Besides the advantage of avoiding any predetermined models which may not
necessarily fit the data, our method registers the diffusion weighted images
(DWIs) and allows model fitting after the registration. This essentially
means that our method can be utilized as a preprocessing step for a wide
assortment of available diffusion models. Our method is also well suited for
clinical applications due to its low computational cost around 5 minutes on
a 2.8GHz Linux machine (without algorithm optimization) to register a pair of
images of typical size 128 x 128 x 80. The main idea involves extraction of
statistical features directly from the diffusion profile, which includes mean
diffusivity, diffusion anisotropy, regional diffusion statistics, and
statistic-map-based edges.