DT-REFinD: diffusion tensor registration with exact finite-strain differential.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Neuroimaging

Background:

• Nonlinear registration of diffusion tensor images (DTIs) is complex due to the need for both reorientation and interpolation strategies.
• Existing methods often approximate gradients, limiting accuracy in tensor image registration.

Purpose of the Study:

• To develop an accurate and efficient diffeomorphic nonlinear registration algorithm for diffusion tensor images.
• To derive and utilize an analytical gradient for the finite-strain reorientation strategy in DTI registration.

Main Methods:

• Proposed the DT-REFinD algorithm incorporating an exact, closed-form gradient derived from computer vision principles.
• Employed velocity field representations of one-parameter subgroups of diffeomorphisms for guaranteed diffeomorphism.
• Compared DT-REFinD against a traditional finite-strain method that neglects reorientation in gradient computation.

Main Results:

• The DT-REFinD algorithm with the exact gradient achieved significantly better registration accuracy compared to the traditional approach.
• Improved alignment was observed across various deformation penalties, interpolation schemes (Euclidean, Log-Euclidean), and dissimilarity measures.
• Performance was validated using metrics including tensor overlap, fractional anisotropy, and inverse consistency.

Conclusions:

• Deriving and utilizing the exact gradient for tensor reorientation is crucial for high-quality DTI registration.
• DT-REFinD offers a robust and accurate solution for diffeomorphic nonlinear registration of diffusion tensor images.
• The algorithm's effectiveness is demonstrated even with alternative reorientation schemes like preservation of principal directions.