Correspondence of closest gradient voxels--a robust registration algorithm

Area of Science:

• Medical Imaging
• Image Processing
• Computational Anatomy

Background:

• Accurate registration of medical imaging volumes is crucial for quantitative analysis and longitudinal studies.
• Existing registration methods can struggle with variations in image intensity, significant patient movement, or incomplete data.

Purpose of the Study:

• To present a robust, automatic volume registration algorithm based on intensity gradients.
• To demonstrate the algorithm's effectiveness under challenging conditions, including unrelated voxel intensities, large displacements, and missing data.
• To provide visualization tools for assessing registration convergence and identifying errors.

Main Methods:

• Developed an automatic registration algorithm utilizing three-dimensional intensity gradients.
• Employed a matching strategy based on iteratively finding correspondences between voxels with high gradient magnitudes.
• Tested the algorithm on T2-weighted and proton-density Magnetic Resonance (MR) volumes with simulated rotations (up to 25 degrees) and translations (up to 25 mm).

Main Results:

• Achieved highly accurate registrations with a mean error of less than one-fifth of a voxel.
• Demonstrated robustness against unrelated inter-volume voxel intensities, significant object displacements, and substantial amounts of missing data.
• Registration times were consistently under 30 minutes per volume pair.

Conclusions:

• The proposed intensity gradient-based algorithm provides a powerful and sequence-independent solution for MR volume registration.
• The algorithm is user-friendly, offering clear visualization of registration convergence and error sources.
• It offers a reliable method for medical image registration, even in the presence of data imperfections.