Collision-constrained deformable image registration framework for discontinuity management

Area of Science:

• Medical Imaging
• Computer Graphics
• Optimization

Background:

• Topological changes like sliding, sources, and sinks pose challenges in image registration.
• Existing methods struggle to manage overlapping objects with independent deformation fields.

Purpose of the Study:

• To propose a general framework for constraining image registration of separate objects.
• To prevent overlapping and maintain registration accuracy using a novel approach.

Main Methods:

• Utilizing the alternating direction method of multipliers (ADMM) for constrained registration.
• Integrating a collision detection algorithm from computer graphics for overlap prevention.
• Employing Free-Form Deformations (FFD) within a divide and conquer optimization strategy.

Main Results:

• The proposed framework successfully prevents intersection between registered objects.
• Demonstrated superior performance in combined intersection prevention and registration accuracy.
• Validated on synthetic data with complex displacements and real-world lung data (DIR-Lab dataset).

Conclusions:

• The ADMM-based framework provides a robust solution for image registration with non-intersection constraints.
• The method generalizes to real medical imaging data, specifically for lung registration.
• This approach enhances the reliability of image registration for complex scenarios.