GRAM: A framework for geodesic registration on anatomical manifolds.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Biomedical Engineering

Background:

• Medical image registration is crucial for analyzing anatomical changes.
• Large deformation registration methods face challenges due to anatomical variability.
• Existing geodesic methods may not reflect biologically plausible anatomical transformations.

Purpose of the Study:

• To develop a novel framework for large deformation medical image registration.
• To address limitations of current geodesic registration methods.
• To leverage learned anatomical variation for more accurate registration.

Main Methods:

• Proposing a framework that decomposes large deformations into series of small deformations.
• Utilizing an empirical manifold representing anatomical variation.
• Applying manifold learning techniques for data visualization and optimal template selection.
• Decomposing large deformations along the shortest path on a learned manifold.

Main Results:

• Demonstrated advantages of the proposed framework over direct registration methods.
• Successfully visualized major data variations using low-dimensional embedding.
• Identified optimal group templates as geodesic means on the manifold.
• Validated the framework using simulated and real brain image databases.

Conclusions:

• The proposed framework offers a more biologically plausible approach to large deformation registration.
• Learning the manifold of anatomical variation improves registration accuracy.
• This method enhances the analysis of medical images with significant anatomical differences.