A LAGRANGIAN FORMULATION FOR STATISTICAL FLUID REGISTRATION.

Area of Science:

• Neuroimaging
• Computational Anatomy
• Statistical Modeling

Background:

• Current brain image registration methods often lack incorporation of empirical statistics on anatomical variability.
• Existing algorithms frequently rely on regularizers ensuring diffeomorphic mappings, potentially overlooking population-specific anatomical data.

Purpose of the Study:

• To develop and evaluate a novel statistical fluid registration method incorporating empirical statistics on population anatomical variability.
• To combine the strengths of large-deformation fluid matching with statistical anatomical data within a Lagrangian framework.

Main Methods:

• Reformulated a Riemannian fluid algorithm using a Lagrangian framework to integrate 0- and 1-order statistics into regularization.
• Applied the non-statistical version (algorithm 0) to 92 2D midline corpus callosum traces from a twin MRI database.
• Computed covariance matrices from distributions and incorporated them separately (algorithms 1 & 2) or together (algorithm 3) into the registration process.

Main Results:

• Generated initial vector fields and deformation tensors using algorithm 0.
• Computed heritability maps.
• Quantified differences in registration power and robustness using vector and tensor-based distances across algorithms 0, 1, 2, and 3.

Conclusions:

• The developed statistical fluid registration method offers a new way to incorporate population variability into anatomical analysis.
• The Lagrangian framework effectively integrates statistical information, potentially improving the accuracy and robustness of brain image registration.
• Further evaluation of heritability maps and distance metrics will elucidate the comparative performance of the statistical integration approaches.